program(1.0)
[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.7.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
{
    func main<ios17>(tensor<int32, [1]> cur_len, tensor<int32, [1, 1]> input_ids, tensor<fp32, [28, 1, 4, 2048, 128]> kv_k, tensor<fp32, [28, 1, 4, 2048, 128]> kv_v) {
            tensor<int32, [2]> var_75 = const()[name = tensor<string, []>("op_75"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [1, 1]> positions = reshape(shape = var_75, x = cur_len)[name = tensor<string, []>("positions")];
            tensor<int32, [1]> var_83_axes_0 = const()[name = tensor<string, []>("op_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<string, []> var_81_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_81_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1]> positions_to_fp16 = cast(dtype = var_81_to_fp16_dtype_0, x = positions)[name = tensor<string, []>("cast_690")];
            tensor<fp16, [1, 1, 1]> var_83_cast_fp16 = expand_dims(axes = var_83_axes_0, x = positions_to_fp16)[name = tensor<string, []>("op_83_cast_fp16")];
            tensor<fp16, [1, 1, 64]> var_88_to_fp16 = const()[name = tensor<string, []>("op_88_to_fp16"), val = tensor<fp16, [1, 1, 64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
            tensor<fp16, [1, 1, 64]> freqs_cast_fp16 = mul(x = var_83_cast_fp16, y = var_88_to_fp16)[name = tensor<string, []>("freqs_cast_fp16")];
            tensor<int32, []> var_91 = const()[name = tensor<string, []>("op_91"), val = tensor<int32, []>(-1)];
            tensor<bool, []> emb_interleave_0 = const()[name = tensor<string, []>("emb_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 1, 128]> emb_cast_fp16 = concat(axis = var_91, interleave = emb_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = tensor<string, []>("emb_cast_fp16")];
            tensor<fp16, [1, 1, 128]> cos_1_cast_fp16 = cos(x = emb_cast_fp16)[name = tensor<string, []>("cos_1_cast_fp16")];
            tensor<fp16, [1, 1, 128]> sin_1_cast_fp16 = sin(x = emb_cast_fp16)[name = tensor<string, []>("sin_1_cast_fp16")];
            tensor<int32, [1, 1, 2048, 1]> pj = const()[name = tensor<string, []>("pj"), val = tensor<int32, [1, 1, 2048, 1]>([[[[0], [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]]]])];
            tensor<int32, [4]> var_105 = const()[name = tensor<string, []>("op_105"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [1, 1, 1, 1]> var_106 = reshape(shape = var_105, x = cur_len)[name = tensor<string, []>("op_106")];
            tensor<bool, [1, 1, 2048, 1]> var_107 = equal(x = pj, y = var_106)[name = tensor<string, []>("op_107")];
            tensor<int32, [1, 1, 1, 2048]> var_118 = const()[name = tensor<string, []>("op_118"), val = tensor<int32, [1, 1, 1, 2048]>([[[[0, 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]]]])];
            tensor<bool, [1, 1, 1, 2048]> attendable = less_equal(x = var_118, y = var_106)[name = tensor<string, []>("attendable")];
            tensor<fp16, [1, 1, 1, 2048]> var_139_after_broadcast_to_fp16 = const()[name = tensor<string, []>("op_139_after_broadcast_to_fp16"), val = tensor<fp16, [1, 1, 1, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256)))];
            tensor<fp16, [1, 1, 1, 2048]> const_0_after_broadcast_to_fp16 = const()[name = tensor<string, []>("const_0_after_broadcast_to_fp16"), val = tensor<fp16, [1, 1, 1, 2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4416)))];
            tensor<fp16, [1, 1, 1, 2048]> attn_mask_cast_fp16 = select(a = var_139_after_broadcast_to_fp16, b = const_0_after_broadcast_to_fp16, cond = attendable)[name = tensor<string, []>("attn_mask_cast_fp16")];
            tensor<int32, []> x_1_batch_dims_0 = const()[name = tensor<string, []>("x_1_batch_dims_0"), val = tensor<int32, []>(0)];
            tensor<bool, []> x_1_validate_indices_0 = const()[name = tensor<string, []>("x_1_validate_indices_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [217232, 512]> tok_weight_to_fp16 = const()[name = tensor<string, []>("tok_weight_to_fp16"), val = tensor<fp16, [217232, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8576)))];
            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
            tensor<int32, []> slice_by_index_0 = const()[name = tensor<string, []>("slice_by_index_0"), val = tensor<int32, []>(217232)];
            tensor<int32, [1, 1]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = tensor<string, []>("add_0")];
            tensor<int32, [1, 1]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
            tensor<int32, []> x_1_cast_fp16_axis_0 = const()[name = tensor<string, []>("x_1_cast_fp16_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [1, 1, 512]> x_1_cast_fp16 = gather(axis = x_1_cast_fp16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0, validate_indices = x_1_validate_indices_0, x = tok_weight_to_fp16)[name = tensor<string, []>("x_1_cast_fp16")];
            tensor<string, []> x_1_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_1_begin_0 = const()[name = tensor<string, []>("k_cache_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_1_end_0 = const()[name = tensor<string, []>("k_cache_1_end_0"), val = tensor<int32, [5]>([1, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_1_end_mask_0 = const()[name = tensor<string, []>("k_cache_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_1_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<string, []> kv_k_to_fp16_dtype_0 = const()[name = tensor<string, []>("kv_k_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [28, 1, 4, 2048, 128]> kv_k_to_fp16 = cast(dtype = kv_k_to_fp16_dtype_0, x = kv_k)[name = tensor<string, []>("cast_688")];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_1_cast_fp16 = slice_by_index(begin = k_cache_1_begin_0, end = k_cache_1_end_0, end_mask = k_cache_1_end_mask_0, squeeze_mask = k_cache_1_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_1_cast_fp16")];
            tensor<int32, [5]> v_cache_1_begin_0 = const()[name = tensor<string, []>("v_cache_1_begin_0"), val = tensor<int32, [5]>([0, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_1_end_0 = const()[name = tensor<string, []>("v_cache_1_end_0"), val = tensor<int32, [5]>([1, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_1_end_mask_0 = const()[name = tensor<string, []>("v_cache_1_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_1_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_1_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<string, []> kv_v_to_fp16_dtype_0 = const()[name = tensor<string, []>("kv_v_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [28, 1, 4, 2048, 128]> kv_v_to_fp16 = cast(dtype = kv_v_to_fp16_dtype_0, x = kv_v)[name = tensor<string, []>("cast_687")];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_1_cast_fp16 = slice_by_index(begin = v_cache_1_begin_0, end = v_cache_1_end_0, end_mask = v_cache_1_end_mask_0, squeeze_mask = v_cache_1_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_1_cast_fp16")];
            tensor<int32, []> var_171 = const()[name = tensor<string, []>("op_171"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_170_promoted = const()[name = tensor<string, []>("op_170_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_1_cast_fp16_to_fp32 = cast(dtype = x_1_cast_fp16_to_fp32_dtype_0, x = x_1_cast_fp16)[name = tensor<string, []>("cast_689")];
            tensor<fp32, [1, 1, 512]> var_180 = pow(x = x_1_cast_fp16_to_fp32, y = var_170_promoted)[name = tensor<string, []>("op_180")];
            tensor<int32, [1]> var_1_axes_0 = const()[name = tensor<string, []>("var_1_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_1_keep_dims_0 = const()[name = tensor<string, []>("var_1_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_1 = reduce_mean(axes = var_1_axes_0, keep_dims = var_1_keep_dims_0, x = var_180)[name = tensor<string, []>("var_1")];
            tensor<string, []> var_1_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_1_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_184_to_fp16 = const()[name = tensor<string, []>("op_184_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_1_to_fp16 = cast(dtype = var_1_to_fp16_dtype_0, x = var_1)[name = tensor<string, []>("cast_686")];
            tensor<fp16, [1, 1, 1]> var_185_cast_fp16 = add(x = var_1_to_fp16, y = var_184_to_fp16)[name = tensor<string, []>("op_185_cast_fp16")];
            tensor<string, []> var_185_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_185_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_186_epsilon_0 = const()[name = tensor<string, []>("op_186_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_185_cast_fp16_to_fp32 = cast(dtype = var_185_cast_fp16_to_fp32_dtype_0, x = var_185_cast_fp16)[name = tensor<string, []>("cast_685")];
            tensor<fp32, [1, 1, 1]> var_186 = rsqrt(epsilon = var_186_epsilon_0, x = var_185_cast_fp16_to_fp32)[name = tensor<string, []>("op_186")];
            tensor<string, []> var_186_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_186_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_186_to_fp16 = cast(dtype = var_186_to_fp16_dtype_0, x = var_186)[name = tensor<string, []>("cast_684")];
            tensor<fp16, [1, 1, 512]> x_7_cast_fp16 = mul(x = x_1_cast_fp16, y = var_186_to_fp16)[name = tensor<string, []>("x_7_cast_fp16")];
            tensor<fp16, [512]> layers_0_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222454208)))];
            tensor<fp16, [1, 1, 512]> x_9_cast_fp16 = mul(x = layers_0_input_layernorm_weight_to_fp16, y = x_7_cast_fp16)[name = tensor<string, []>("x_9_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222455296)))];
            tensor<fp16, [1536]> linear_0_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224028224)))];
            tensor<fp16, [1, 1, 1536]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = x_9_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
            tensor<int32, [4]> var_200 = const()[name = tensor<string, []>("op_200"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_11_cast_fp16 = reshape(shape = var_200, x = linear_0_cast_fp16)[name = tensor<string, []>("x_11_cast_fp16")];
            tensor<string, []> x_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_11_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_170_promoted_1 = const()[name = tensor<string, []>("op_170_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_11_cast_fp16_to_fp32 = cast(dtype = x_11_cast_fp16_to_fp32_dtype_0, x = x_11_cast_fp16)[name = tensor<string, []>("cast_683")];
            tensor<fp32, [1, 1, 12, 128]> var_204 = pow(x = x_11_cast_fp16_to_fp32, y = var_170_promoted_1)[name = tensor<string, []>("op_204")];
            tensor<int32, [1]> var_3_axes_0 = const()[name = tensor<string, []>("var_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_3_keep_dims_0 = const()[name = tensor<string, []>("var_3_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_3 = reduce_mean(axes = var_3_axes_0, keep_dims = var_3_keep_dims_0, x = var_204)[name = tensor<string, []>("var_3")];
            tensor<string, []> var_3_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_3_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_208_to_fp16 = const()[name = tensor<string, []>("op_208_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_3_to_fp16 = cast(dtype = var_3_to_fp16_dtype_0, x = var_3)[name = tensor<string, []>("cast_682")];
            tensor<fp16, [1, 1, 12, 1]> var_209_cast_fp16 = add(x = var_3_to_fp16, y = var_208_to_fp16)[name = tensor<string, []>("op_209_cast_fp16")];
            tensor<string, []> var_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_209_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_210_epsilon_0 = const()[name = tensor<string, []>("op_210_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_209_cast_fp16_to_fp32 = cast(dtype = var_209_cast_fp16_to_fp32_dtype_0, x = var_209_cast_fp16)[name = tensor<string, []>("cast_681")];
            tensor<fp32, [1, 1, 12, 1]> var_210 = rsqrt(epsilon = var_210_epsilon_0, x = var_209_cast_fp16_to_fp32)[name = tensor<string, []>("op_210")];
            tensor<string, []> var_210_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_210_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_210_to_fp16 = cast(dtype = var_210_to_fp16_dtype_0, x = var_210)[name = tensor<string, []>("cast_680")];
            tensor<fp16, [1, 1, 12, 128]> x_17_cast_fp16 = mul(x = x_11_cast_fp16, y = var_210_to_fp16)[name = tensor<string, []>("x_17_cast_fp16")];
            tensor<fp16, [128]> layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224031360)))];
            tensor<fp16, [1, 1, 12, 128]> var_212_cast_fp16 = mul(x = layers_0_self_attn_q_norm_weight_to_fp16, y = x_17_cast_fp16)[name = tensor<string, []>("op_212_cast_fp16")];
            tensor<int32, [4]> q_1_perm_0 = const()[name = tensor<string, []>("q_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224031680)))];
            tensor<fp16, [512]> linear_1_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_1_bias_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224556032)))];
            tensor<fp16, [1, 1, 512]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = x_9_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
            tensor<int32, [4]> var_216 = const()[name = tensor<string, []>("op_216"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_19_cast_fp16 = reshape(shape = var_216, x = linear_1_cast_fp16)[name = tensor<string, []>("x_19_cast_fp16")];
            tensor<string, []> x_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_19_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_170_promoted_2 = const()[name = tensor<string, []>("op_170_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_19_cast_fp16_to_fp32 = cast(dtype = x_19_cast_fp16_to_fp32_dtype_0, x = x_19_cast_fp16)[name = tensor<string, []>("cast_679")];
            tensor<fp32, [1, 1, 4, 128]> var_220 = pow(x = x_19_cast_fp16_to_fp32, y = var_170_promoted_2)[name = tensor<string, []>("op_220")];
            tensor<int32, [1]> var_5_axes_0 = const()[name = tensor<string, []>("var_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_5_keep_dims_0 = const()[name = tensor<string, []>("var_5_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_5 = reduce_mean(axes = var_5_axes_0, keep_dims = var_5_keep_dims_0, x = var_220)[name = tensor<string, []>("var_5")];
            tensor<string, []> var_5_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_5_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_224_to_fp16 = const()[name = tensor<string, []>("op_224_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_5_to_fp16 = cast(dtype = var_5_to_fp16_dtype_0, x = var_5)[name = tensor<string, []>("cast_678")];
            tensor<fp16, [1, 1, 4, 1]> var_225_cast_fp16 = add(x = var_5_to_fp16, y = var_224_to_fp16)[name = tensor<string, []>("op_225_cast_fp16")];
            tensor<string, []> var_225_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_225_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_226_epsilon_0 = const()[name = tensor<string, []>("op_226_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_225_cast_fp16_to_fp32 = cast(dtype = var_225_cast_fp16_to_fp32_dtype_0, x = var_225_cast_fp16)[name = tensor<string, []>("cast_677")];
            tensor<fp32, [1, 1, 4, 1]> var_226 = rsqrt(epsilon = var_226_epsilon_0, x = var_225_cast_fp16_to_fp32)[name = tensor<string, []>("op_226")];
            tensor<string, []> var_226_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_226_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_226_to_fp16 = cast(dtype = var_226_to_fp16_dtype_0, x = var_226)[name = tensor<string, []>("cast_676")];
            tensor<fp16, [1, 1, 4, 128]> x_25_cast_fp16 = mul(x = x_19_cast_fp16, y = var_226_to_fp16)[name = tensor<string, []>("x_25_cast_fp16")];
            tensor<fp16, [128]> layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224557120)))];
            tensor<fp16, [1, 1, 4, 128]> var_228_cast_fp16 = mul(x = layers_0_self_attn_k_norm_weight_to_fp16, y = x_25_cast_fp16)[name = tensor<string, []>("op_228_cast_fp16")];
            tensor<int32, [4]> k_1_perm_0 = const()[name = tensor<string, []>("k_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(224557440)))];
            tensor<fp16, [1, 1, 512]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = x_9_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
            tensor<int32, [4]> var_232 = const()[name = tensor<string, []>("op_232"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_233_cast_fp16 = reshape(shape = var_232, x = linear_2_cast_fp16)[name = tensor<string, []>("op_233_cast_fp16")];
            tensor<int32, [4]> v_1_perm_0 = const()[name = tensor<string, []>("v_1_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [1]> cos_3_axes_0 = const()[name = tensor<string, []>("cos_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 128]> cos_3_cast_fp16 = expand_dims(axes = cos_3_axes_0, x = cos_1_cast_fp16)[name = tensor<string, []>("cos_3_cast_fp16")];
            tensor<int32, [1]> sin_3_axes_0 = const()[name = tensor<string, []>("sin_3_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1, 128]> sin_3_cast_fp16 = expand_dims(axes = sin_3_axes_0, x = sin_1_cast_fp16)[name = tensor<string, []>("sin_3_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_1_cast_fp16 = transpose(perm = q_1_perm_0, x = var_212_cast_fp16)[name = tensor<string, []>("transpose_111")];
            tensor<fp16, [1, 12, 1, 128]> var_237_cast_fp16 = mul(x = q_1_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_237_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = tensor<string, []>("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = tensor<string, []>("x1_1_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = tensor<string, []>("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_1_cast_fp16)[name = tensor<string, []>("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = tensor<string, []>("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = tensor<string, []>("x2_1_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = tensor<string, []>("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_1_cast_fp16)[name = tensor<string, []>("x2_1_cast_fp16")];
            tensor<fp16, []> const_1_promoted_to_fp16 = const()[name = tensor<string, []>("const_1_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_240_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_1_promoted_to_fp16)[name = tensor<string, []>("op_240_cast_fp16")];
            tensor<bool, []> var_242_interleave_0 = const()[name = tensor<string, []>("op_242_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_242_cast_fp16 = concat(axis = var_171, interleave = var_242_interleave_0, values = (var_240_cast_fp16, x1_1_cast_fp16))[name = tensor<string, []>("op_242_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_243_cast_fp16 = mul(x = var_242_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_243_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_3_cast_fp16 = add(x = var_237_cast_fp16, y = var_243_cast_fp16)[name = tensor<string, []>("q_3_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_1_cast_fp16 = transpose(perm = k_1_perm_0, x = var_228_cast_fp16)[name = tensor<string, []>("transpose_110")];
            tensor<fp16, [1, 4, 1, 128]> var_245_cast_fp16 = mul(x = k_1_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_245_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = tensor<string, []>("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = tensor<string, []>("x1_3_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = tensor<string, []>("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_1_cast_fp16)[name = tensor<string, []>("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = tensor<string, []>("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = tensor<string, []>("x2_3_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = tensor<string, []>("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_1_cast_fp16)[name = tensor<string, []>("x2_3_cast_fp16")];
            tensor<fp16, []> const_2_promoted_to_fp16 = const()[name = tensor<string, []>("const_2_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_248_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_2_promoted_to_fp16)[name = tensor<string, []>("op_248_cast_fp16")];
            tensor<bool, []> var_250_interleave_0 = const()[name = tensor<string, []>("op_250_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_250_cast_fp16 = concat(axis = var_171, interleave = var_250_interleave_0, values = (var_248_cast_fp16, x1_3_cast_fp16))[name = tensor<string, []>("op_250_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_251_cast_fp16 = mul(x = var_250_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_251_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_3_cast_fp16 = add(x = var_245_cast_fp16, y = var_251_cast_fp16)[name = tensor<string, []>("k_3_cast_fp16")];
            tensor<fp16, []> var_163_to_fp16 = const()[name = tensor<string, []>("op_163_to_fp16"), val = tensor<fp16, []>(0x1p+0)];
            tensor<string, []> update_mask_to_fp16_dtype_0 = const()[name = tensor<string, []>("update_mask_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 2048, 1]> var_107_to_fp16 = cast(dtype = update_mask_to_fp16_dtype_0, x = var_107)[name = tensor<string, []>("cast_675")];
            tensor<fp16, [1, 1, 2048, 1]> var_253_cast_fp16 = sub(x = var_163_to_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_253_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_254_cast_fp16 = mul(x = k_cache_1_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_254_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_255_cast_fp16 = mul(x = k_3_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_255_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_1_cast_fp16 = add(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = tensor<string, []>("k_full_1_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_258_cast_fp16 = mul(x = v_cache_1_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_258_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_233_cast_fp16)[name = tensor<string, []>("transpose_109")];
            tensor<fp16, [1, 4, 2048, 128]> var_259_cast_fp16 = mul(x = v_1_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_259_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_1_cast_fp16 = add(x = var_258_cast_fp16, y = var_259_cast_fp16)[name = tensor<string, []>("v_full_1_cast_fp16")];
            tensor<int32, [1]> var_261_axes_0 = const()[name = tensor<string, []>("op_261_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_261_cast_fp16 = expand_dims(axes = var_261_axes_0, x = k_full_1_cast_fp16)[name = tensor<string, []>("op_261_cast_fp16")];
            tensor<int32, [5]> var_263_reps_0 = const()[name = tensor<string, []>("op_263_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_263_cast_fp16 = tile(reps = var_263_reps_0, x = var_261_cast_fp16)[name = tensor<string, []>("op_263_cast_fp16")];
            tensor<int32, [4]> var_264 = const()[name = tensor<string, []>("op_264"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_1_cast_fp16 = reshape(shape = var_264, x = var_263_cast_fp16)[name = tensor<string, []>("k_rep_1_cast_fp16")];
            tensor<int32, [1]> var_266_axes_0 = const()[name = tensor<string, []>("op_266_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_266_cast_fp16 = expand_dims(axes = var_266_axes_0, x = v_full_1_cast_fp16)[name = tensor<string, []>("op_266_cast_fp16")];
            tensor<int32, [5]> var_268_reps_0 = const()[name = tensor<string, []>("op_268_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_268_cast_fp16 = tile(reps = var_268_reps_0, x = var_266_cast_fp16)[name = tensor<string, []>("op_268_cast_fp16")];
            tensor<int32, [4]> var_269 = const()[name = tensor<string, []>("op_269"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_1_cast_fp16 = reshape(shape = var_269, x = var_268_cast_fp16)[name = tensor<string, []>("v_rep_1_cast_fp16")];
            tensor<bool, []> var_272_transpose_x_1 = const()[name = tensor<string, []>("op_272_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_272_transpose_y_1 = const()[name = tensor<string, []>("op_272_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_272_cast_fp16 = matmul(transpose_x = var_272_transpose_x_1, transpose_y = var_272_transpose_y_1, x = q_3_cast_fp16, y = k_rep_1_cast_fp16)[name = tensor<string, []>("op_272_cast_fp16")];
            tensor<fp16, []> var_273_to_fp16 = const()[name = tensor<string, []>("op_273_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_1_cast_fp16 = mul(x = var_272_cast_fp16, y = var_273_to_fp16)[name = tensor<string, []>("attn_1_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_3_cast_fp16 = add(x = attn_1_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
            tensor<string, []> input_3_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_3_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_3_cast_fp16_to_fp32 = cast(dtype = input_3_cast_fp16_to_fp32_dtype_0, x = input_3_cast_fp16)[name = tensor<string, []>("cast_674")];
            tensor<fp32, [1, 12, 1, 2048]> attn_3 = softmax(axis = var_171, x = input_3_cast_fp16_to_fp32)[name = tensor<string, []>("attn_3")];
            tensor<bool, []> out_1_transpose_x_0 = const()[name = tensor<string, []>("out_1_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_1_transpose_y_0 = const()[name = tensor<string, []>("out_1_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_3_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_3_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_3_to_fp16 = cast(dtype = attn_3_to_fp16_dtype_0, x = attn_3)[name = tensor<string, []>("cast_673")];
            tensor<fp16, [1, 12, 1, 128]> out_1_cast_fp16 = matmul(transpose_x = out_1_transpose_x_0, transpose_y = out_1_transpose_y_0, x = attn_3_to_fp16, y = v_rep_1_cast_fp16)[name = tensor<string, []>("out_1_cast_fp16")];
            tensor<int32, [4]> var_278_perm_0 = const()[name = tensor<string, []>("op_278_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_280 = const()[name = tensor<string, []>("op_280"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_278_cast_fp16 = transpose(perm = var_278_perm_0, x = out_1_cast_fp16)[name = tensor<string, []>("transpose_108")];
            tensor<fp16, [1, 1, 1536]> x_27_cast_fp16 = reshape(shape = var_280, x = var_278_cast_fp16)[name = tensor<string, []>("x_27_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(225081792)))];
            tensor<fp16, [1, 1, 512]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = x_27_cast_fp16)[name = tensor<string, []>("linear_3_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_29_cast_fp16 = add(x = x_1_cast_fp16, y = linear_3_cast_fp16)[name = tensor<string, []>("x_29_cast_fp16")];
            tensor<string, []> x_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_29_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_170_promoted_3 = const()[name = tensor<string, []>("op_170_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_29_cast_fp16_to_fp32 = cast(dtype = x_29_cast_fp16_to_fp32_dtype_0, x = x_29_cast_fp16)[name = tensor<string, []>("cast_672")];
            tensor<fp32, [1, 1, 512]> var_291 = pow(x = x_29_cast_fp16_to_fp32, y = var_170_promoted_3)[name = tensor<string, []>("op_291")];
            tensor<int32, [1]> var_7_axes_0 = const()[name = tensor<string, []>("var_7_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_7_keep_dims_0 = const()[name = tensor<string, []>("var_7_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_7 = reduce_mean(axes = var_7_axes_0, keep_dims = var_7_keep_dims_0, x = var_291)[name = tensor<string, []>("var_7")];
            tensor<string, []> var_7_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_7_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_295_to_fp16 = const()[name = tensor<string, []>("op_295_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_7_to_fp16 = cast(dtype = var_7_to_fp16_dtype_0, x = var_7)[name = tensor<string, []>("cast_671")];
            tensor<fp16, [1, 1, 1]> var_296_cast_fp16 = add(x = var_7_to_fp16, y = var_295_to_fp16)[name = tensor<string, []>("op_296_cast_fp16")];
            tensor<string, []> var_296_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_296_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_297_epsilon_0 = const()[name = tensor<string, []>("op_297_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_296_cast_fp16_to_fp32 = cast(dtype = var_296_cast_fp16_to_fp32_dtype_0, x = var_296_cast_fp16)[name = tensor<string, []>("cast_670")];
            tensor<fp32, [1, 1, 1]> var_297 = rsqrt(epsilon = var_297_epsilon_0, x = var_296_cast_fp16_to_fp32)[name = tensor<string, []>("op_297")];
            tensor<string, []> var_297_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_297_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_297_to_fp16 = cast(dtype = var_297_to_fp16_dtype_0, x = var_297)[name = tensor<string, []>("cast_669")];
            tensor<fp16, [1, 1, 512]> x_35_cast_fp16 = mul(x = x_29_cast_fp16, y = var_297_to_fp16)[name = tensor<string, []>("x_35_cast_fp16")];
            tensor<fp16, [512]> layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226654720)))];
            tensor<fp16, [1, 1, 512]> x_37_cast_fp16 = mul(x = layers_0_post_attention_layernorm_weight_to_fp16, y = x_35_cast_fp16)[name = tensor<string, []>("x_37_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(226655808)))];
            tensor<fp16, [1, 1, 1536]> linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_mlp_gate_proj_weight_to_fp16, x = x_37_cast_fp16)[name = tensor<string, []>("linear_4_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_308_cast_fp16 = silu(x = linear_4_cast_fp16)[name = tensor<string, []>("op_308_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(228228736)))];
            tensor<fp16, [1, 1, 1536]> linear_5_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_mlp_up_proj_weight_to_fp16, x = x_37_cast_fp16)[name = tensor<string, []>("linear_5_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_39_cast_fp16 = mul(x = var_308_cast_fp16, y = linear_5_cast_fp16)[name = tensor<string, []>("x_39_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(229801664)))];
            tensor<fp16, [1, 1, 512]> linear_6_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_0_mlp_down_proj_weight_to_fp16, x = x_39_cast_fp16)[name = tensor<string, []>("linear_6_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_41_cast_fp16 = add(x = x_29_cast_fp16, y = linear_6_cast_fp16)[name = tensor<string, []>("x_41_cast_fp16")];
            tensor<string, []> x_41_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_41_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_3_begin_0 = const()[name = tensor<string, []>("k_cache_3_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_3_end_0 = const()[name = tensor<string, []>("k_cache_3_end_0"), val = tensor<int32, [5]>([2, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_3_end_mask_0 = const()[name = tensor<string, []>("k_cache_3_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_3_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_3_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_3_cast_fp16 = slice_by_index(begin = k_cache_3_begin_0, end = k_cache_3_end_0, end_mask = k_cache_3_end_mask_0, squeeze_mask = k_cache_3_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_3_cast_fp16")];
            tensor<int32, [5]> v_cache_3_begin_0 = const()[name = tensor<string, []>("v_cache_3_begin_0"), val = tensor<int32, [5]>([1, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_3_end_0 = const()[name = tensor<string, []>("v_cache_3_end_0"), val = tensor<int32, [5]>([2, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_3_end_mask_0 = const()[name = tensor<string, []>("v_cache_3_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_3_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_3_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_3_cast_fp16 = slice_by_index(begin = v_cache_3_begin_0, end = v_cache_3_end_0, end_mask = v_cache_3_end_mask_0, squeeze_mask = v_cache_3_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_3_cast_fp16")];
            tensor<int32, []> var_340 = const()[name = tensor<string, []>("op_340"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_339_promoted = const()[name = tensor<string, []>("op_339_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_41_cast_fp16_to_fp32 = cast(dtype = x_41_cast_fp16_to_fp32_dtype_0, x = x_41_cast_fp16)[name = tensor<string, []>("cast_668")];
            tensor<fp32, [1, 1, 512]> var_349 = pow(x = x_41_cast_fp16_to_fp32, y = var_339_promoted)[name = tensor<string, []>("op_349")];
            tensor<int32, [1]> var_9_axes_0 = const()[name = tensor<string, []>("var_9_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_9_keep_dims_0 = const()[name = tensor<string, []>("var_9_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_9 = reduce_mean(axes = var_9_axes_0, keep_dims = var_9_keep_dims_0, x = var_349)[name = tensor<string, []>("var_9")];
            tensor<string, []> var_9_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_9_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_353_to_fp16 = const()[name = tensor<string, []>("op_353_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_9_to_fp16 = cast(dtype = var_9_to_fp16_dtype_0, x = var_9)[name = tensor<string, []>("cast_667")];
            tensor<fp16, [1, 1, 1]> var_354_cast_fp16 = add(x = var_9_to_fp16, y = var_353_to_fp16)[name = tensor<string, []>("op_354_cast_fp16")];
            tensor<string, []> var_354_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_354_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_355_epsilon_0 = const()[name = tensor<string, []>("op_355_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_354_cast_fp16_to_fp32 = cast(dtype = var_354_cast_fp16_to_fp32_dtype_0, x = var_354_cast_fp16)[name = tensor<string, []>("cast_666")];
            tensor<fp32, [1, 1, 1]> var_355 = rsqrt(epsilon = var_355_epsilon_0, x = var_354_cast_fp16_to_fp32)[name = tensor<string, []>("op_355")];
            tensor<string, []> var_355_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_355_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_355_to_fp16 = cast(dtype = var_355_to_fp16_dtype_0, x = var_355)[name = tensor<string, []>("cast_665")];
            tensor<fp16, [1, 1, 512]> x_47_cast_fp16 = mul(x = x_41_cast_fp16, y = var_355_to_fp16)[name = tensor<string, []>("x_47_cast_fp16")];
            tensor<fp16, [512]> layers_1_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231374592)))];
            tensor<fp16, [1, 1, 512]> x_49_cast_fp16 = mul(x = layers_1_input_layernorm_weight_to_fp16, y = x_47_cast_fp16)[name = tensor<string, []>("x_49_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231375680)))];
            tensor<fp16, [1, 1, 1536]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = x_49_cast_fp16)[name = tensor<string, []>("linear_7_cast_fp16")];
            tensor<int32, [4]> var_369 = const()[name = tensor<string, []>("op_369"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_51_cast_fp16 = reshape(shape = var_369, x = linear_7_cast_fp16)[name = tensor<string, []>("x_51_cast_fp16")];
            tensor<string, []> x_51_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_51_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_339_promoted_1 = const()[name = tensor<string, []>("op_339_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_51_cast_fp16_to_fp32 = cast(dtype = x_51_cast_fp16_to_fp32_dtype_0, x = x_51_cast_fp16)[name = tensor<string, []>("cast_664")];
            tensor<fp32, [1, 1, 12, 128]> var_373 = pow(x = x_51_cast_fp16_to_fp32, y = var_339_promoted_1)[name = tensor<string, []>("op_373")];
            tensor<int32, [1]> var_11_axes_0 = const()[name = tensor<string, []>("var_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_11_keep_dims_0 = const()[name = tensor<string, []>("var_11_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_11 = reduce_mean(axes = var_11_axes_0, keep_dims = var_11_keep_dims_0, x = var_373)[name = tensor<string, []>("var_11")];
            tensor<string, []> var_11_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_11_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_377_to_fp16 = const()[name = tensor<string, []>("op_377_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_11_to_fp16 = cast(dtype = var_11_to_fp16_dtype_0, x = var_11)[name = tensor<string, []>("cast_663")];
            tensor<fp16, [1, 1, 12, 1]> var_378_cast_fp16 = add(x = var_11_to_fp16, y = var_377_to_fp16)[name = tensor<string, []>("op_378_cast_fp16")];
            tensor<string, []> var_378_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_378_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_379_epsilon_0 = const()[name = tensor<string, []>("op_379_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_378_cast_fp16_to_fp32 = cast(dtype = var_378_cast_fp16_to_fp32_dtype_0, x = var_378_cast_fp16)[name = tensor<string, []>("cast_662")];
            tensor<fp32, [1, 1, 12, 1]> var_379 = rsqrt(epsilon = var_379_epsilon_0, x = var_378_cast_fp16_to_fp32)[name = tensor<string, []>("op_379")];
            tensor<string, []> var_379_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_379_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_379_to_fp16 = cast(dtype = var_379_to_fp16_dtype_0, x = var_379)[name = tensor<string, []>("cast_661")];
            tensor<fp16, [1, 1, 12, 128]> x_57_cast_fp16 = mul(x = x_51_cast_fp16, y = var_379_to_fp16)[name = tensor<string, []>("x_57_cast_fp16")];
            tensor<fp16, [128]> layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232948608)))];
            tensor<fp16, [1, 1, 12, 128]> var_381_cast_fp16 = mul(x = layers_1_self_attn_q_norm_weight_to_fp16, y = x_57_cast_fp16)[name = tensor<string, []>("op_381_cast_fp16")];
            tensor<int32, [4]> q_5_perm_0 = const()[name = tensor<string, []>("q_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232948928)))];
            tensor<fp16, [1, 1, 512]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = x_49_cast_fp16)[name = tensor<string, []>("linear_8_cast_fp16")];
            tensor<int32, [4]> var_385 = const()[name = tensor<string, []>("op_385"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_59_cast_fp16 = reshape(shape = var_385, x = linear_8_cast_fp16)[name = tensor<string, []>("x_59_cast_fp16")];
            tensor<string, []> x_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_59_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_339_promoted_2 = const()[name = tensor<string, []>("op_339_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_59_cast_fp16_to_fp32 = cast(dtype = x_59_cast_fp16_to_fp32_dtype_0, x = x_59_cast_fp16)[name = tensor<string, []>("cast_660")];
            tensor<fp32, [1, 1, 4, 128]> var_389 = pow(x = x_59_cast_fp16_to_fp32, y = var_339_promoted_2)[name = tensor<string, []>("op_389")];
            tensor<int32, [1]> var_13_axes_0 = const()[name = tensor<string, []>("var_13_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_13_keep_dims_0 = const()[name = tensor<string, []>("var_13_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_13 = reduce_mean(axes = var_13_axes_0, keep_dims = var_13_keep_dims_0, x = var_389)[name = tensor<string, []>("var_13")];
            tensor<string, []> var_13_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_13_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_393_to_fp16 = const()[name = tensor<string, []>("op_393_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_13_to_fp16 = cast(dtype = var_13_to_fp16_dtype_0, x = var_13)[name = tensor<string, []>("cast_659")];
            tensor<fp16, [1, 1, 4, 1]> var_394_cast_fp16 = add(x = var_13_to_fp16, y = var_393_to_fp16)[name = tensor<string, []>("op_394_cast_fp16")];
            tensor<string, []> var_394_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_394_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_395_epsilon_0 = const()[name = tensor<string, []>("op_395_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_394_cast_fp16_to_fp32 = cast(dtype = var_394_cast_fp16_to_fp32_dtype_0, x = var_394_cast_fp16)[name = tensor<string, []>("cast_658")];
            tensor<fp32, [1, 1, 4, 1]> var_395 = rsqrt(epsilon = var_395_epsilon_0, x = var_394_cast_fp16_to_fp32)[name = tensor<string, []>("op_395")];
            tensor<string, []> var_395_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_395_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_395_to_fp16 = cast(dtype = var_395_to_fp16_dtype_0, x = var_395)[name = tensor<string, []>("cast_657")];
            tensor<fp16, [1, 1, 4, 128]> x_65_cast_fp16 = mul(x = x_59_cast_fp16, y = var_395_to_fp16)[name = tensor<string, []>("x_65_cast_fp16")];
            tensor<fp16, [128]> layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233473280)))];
            tensor<fp16, [1, 1, 4, 128]> var_397_cast_fp16 = mul(x = layers_1_self_attn_k_norm_weight_to_fp16, y = x_65_cast_fp16)[name = tensor<string, []>("op_397_cast_fp16")];
            tensor<int32, [4]> k_5_perm_0 = const()[name = tensor<string, []>("k_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233473600)))];
            tensor<fp16, [1, 1, 512]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = x_49_cast_fp16)[name = tensor<string, []>("linear_9_cast_fp16")];
            tensor<int32, [4]> var_401 = const()[name = tensor<string, []>("op_401"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_402_cast_fp16 = reshape(shape = var_401, x = linear_9_cast_fp16)[name = tensor<string, []>("op_402_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = tensor<string, []>("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_5_cast_fp16 = transpose(perm = q_5_perm_0, x = var_381_cast_fp16)[name = tensor<string, []>("transpose_107")];
            tensor<fp16, [1, 12, 1, 128]> var_406_cast_fp16 = mul(x = q_5_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_406_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = tensor<string, []>("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = tensor<string, []>("x1_5_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = tensor<string, []>("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_5_cast_fp16)[name = tensor<string, []>("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = tensor<string, []>("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = tensor<string, []>("x2_5_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = tensor<string, []>("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_5_cast_fp16)[name = tensor<string, []>("x2_5_cast_fp16")];
            tensor<fp16, []> const_3_promoted_to_fp16 = const()[name = tensor<string, []>("const_3_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_409_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_3_promoted_to_fp16)[name = tensor<string, []>("op_409_cast_fp16")];
            tensor<bool, []> var_411_interleave_0 = const()[name = tensor<string, []>("op_411_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_411_cast_fp16 = concat(axis = var_340, interleave = var_411_interleave_0, values = (var_409_cast_fp16, x1_5_cast_fp16))[name = tensor<string, []>("op_411_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_412_cast_fp16 = mul(x = var_411_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_412_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_7_cast_fp16 = add(x = var_406_cast_fp16, y = var_412_cast_fp16)[name = tensor<string, []>("q_7_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_5_cast_fp16 = transpose(perm = k_5_perm_0, x = var_397_cast_fp16)[name = tensor<string, []>("transpose_106")];
            tensor<fp16, [1, 4, 1, 128]> var_414_cast_fp16 = mul(x = k_5_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_414_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = tensor<string, []>("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = tensor<string, []>("x1_7_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = tensor<string, []>("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_5_cast_fp16)[name = tensor<string, []>("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = tensor<string, []>("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = tensor<string, []>("x2_7_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = tensor<string, []>("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_5_cast_fp16)[name = tensor<string, []>("x2_7_cast_fp16")];
            tensor<fp16, []> const_4_promoted_to_fp16 = const()[name = tensor<string, []>("const_4_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_417_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_4_promoted_to_fp16)[name = tensor<string, []>("op_417_cast_fp16")];
            tensor<bool, []> var_419_interleave_0 = const()[name = tensor<string, []>("op_419_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_419_cast_fp16 = concat(axis = var_340, interleave = var_419_interleave_0, values = (var_417_cast_fp16, x1_7_cast_fp16))[name = tensor<string, []>("op_419_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_420_cast_fp16 = mul(x = var_419_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_420_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_7_cast_fp16 = add(x = var_414_cast_fp16, y = var_420_cast_fp16)[name = tensor<string, []>("k_7_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_423_cast_fp16 = mul(x = k_cache_3_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_423_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_424_cast_fp16 = mul(x = k_7_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_424_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_3_cast_fp16 = add(x = var_423_cast_fp16, y = var_424_cast_fp16)[name = tensor<string, []>("k_full_3_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_427_cast_fp16 = mul(x = v_cache_3_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_427_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_402_cast_fp16)[name = tensor<string, []>("transpose_105")];
            tensor<fp16, [1, 4, 2048, 128]> var_428_cast_fp16 = mul(x = v_3_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_428_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_3_cast_fp16 = add(x = var_427_cast_fp16, y = var_428_cast_fp16)[name = tensor<string, []>("v_full_3_cast_fp16")];
            tensor<int32, [1]> var_430_axes_0 = const()[name = tensor<string, []>("op_430_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_430_cast_fp16 = expand_dims(axes = var_430_axes_0, x = k_full_3_cast_fp16)[name = tensor<string, []>("op_430_cast_fp16")];
            tensor<int32, [5]> var_432_reps_0 = const()[name = tensor<string, []>("op_432_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_432_cast_fp16 = tile(reps = var_432_reps_0, x = var_430_cast_fp16)[name = tensor<string, []>("op_432_cast_fp16")];
            tensor<int32, [4]> var_433 = const()[name = tensor<string, []>("op_433"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_3_cast_fp16 = reshape(shape = var_433, x = var_432_cast_fp16)[name = tensor<string, []>("k_rep_3_cast_fp16")];
            tensor<int32, [1]> var_435_axes_0 = const()[name = tensor<string, []>("op_435_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_435_cast_fp16 = expand_dims(axes = var_435_axes_0, x = v_full_3_cast_fp16)[name = tensor<string, []>("op_435_cast_fp16")];
            tensor<int32, [5]> var_437_reps_0 = const()[name = tensor<string, []>("op_437_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_437_cast_fp16 = tile(reps = var_437_reps_0, x = var_435_cast_fp16)[name = tensor<string, []>("op_437_cast_fp16")];
            tensor<int32, [4]> var_438 = const()[name = tensor<string, []>("op_438"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_3_cast_fp16 = reshape(shape = var_438, x = var_437_cast_fp16)[name = tensor<string, []>("v_rep_3_cast_fp16")];
            tensor<bool, []> var_441_transpose_x_1 = const()[name = tensor<string, []>("op_441_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_441_transpose_y_1 = const()[name = tensor<string, []>("op_441_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_441_cast_fp16 = matmul(transpose_x = var_441_transpose_x_1, transpose_y = var_441_transpose_y_1, x = q_7_cast_fp16, y = k_rep_3_cast_fp16)[name = tensor<string, []>("op_441_cast_fp16")];
            tensor<fp16, []> var_442_to_fp16 = const()[name = tensor<string, []>("op_442_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_5_cast_fp16 = mul(x = var_441_cast_fp16, y = var_442_to_fp16)[name = tensor<string, []>("attn_5_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_7_cast_fp16 = add(x = attn_5_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
            tensor<string, []> input_7_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_7_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_7_cast_fp16_to_fp32 = cast(dtype = input_7_cast_fp16_to_fp32_dtype_0, x = input_7_cast_fp16)[name = tensor<string, []>("cast_656")];
            tensor<fp32, [1, 12, 1, 2048]> attn_7 = softmax(axis = var_340, x = input_7_cast_fp16_to_fp32)[name = tensor<string, []>("attn_7")];
            tensor<bool, []> out_3_transpose_x_0 = const()[name = tensor<string, []>("out_3_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_3_transpose_y_0 = const()[name = tensor<string, []>("out_3_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_7_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_7_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_7_to_fp16 = cast(dtype = attn_7_to_fp16_dtype_0, x = attn_7)[name = tensor<string, []>("cast_655")];
            tensor<fp16, [1, 12, 1, 128]> out_3_cast_fp16 = matmul(transpose_x = out_3_transpose_x_0, transpose_y = out_3_transpose_y_0, x = attn_7_to_fp16, y = v_rep_3_cast_fp16)[name = tensor<string, []>("out_3_cast_fp16")];
            tensor<int32, [4]> var_447_perm_0 = const()[name = tensor<string, []>("op_447_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_449 = const()[name = tensor<string, []>("op_449"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_447_cast_fp16 = transpose(perm = var_447_perm_0, x = out_3_cast_fp16)[name = tensor<string, []>("transpose_104")];
            tensor<fp16, [1, 1, 1536]> x_67_cast_fp16 = reshape(shape = var_449, x = var_447_cast_fp16)[name = tensor<string, []>("x_67_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233997952)))];
            tensor<fp16, [1, 1, 512]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = x_67_cast_fp16)[name = tensor<string, []>("linear_10_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_69_cast_fp16 = add(x = x_41_cast_fp16, y = linear_10_cast_fp16)[name = tensor<string, []>("x_69_cast_fp16")];
            tensor<string, []> x_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_69_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_339_promoted_3 = const()[name = tensor<string, []>("op_339_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_69_cast_fp16_to_fp32 = cast(dtype = x_69_cast_fp16_to_fp32_dtype_0, x = x_69_cast_fp16)[name = tensor<string, []>("cast_654")];
            tensor<fp32, [1, 1, 512]> var_460 = pow(x = x_69_cast_fp16_to_fp32, y = var_339_promoted_3)[name = tensor<string, []>("op_460")];
            tensor<int32, [1]> var_15_axes_0 = const()[name = tensor<string, []>("var_15_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_15_keep_dims_0 = const()[name = tensor<string, []>("var_15_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_15 = reduce_mean(axes = var_15_axes_0, keep_dims = var_15_keep_dims_0, x = var_460)[name = tensor<string, []>("var_15")];
            tensor<string, []> var_15_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_15_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_464_to_fp16 = const()[name = tensor<string, []>("op_464_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_15_to_fp16 = cast(dtype = var_15_to_fp16_dtype_0, x = var_15)[name = tensor<string, []>("cast_653")];
            tensor<fp16, [1, 1, 1]> var_465_cast_fp16 = add(x = var_15_to_fp16, y = var_464_to_fp16)[name = tensor<string, []>("op_465_cast_fp16")];
            tensor<string, []> var_465_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_465_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_466_epsilon_0 = const()[name = tensor<string, []>("op_466_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_465_cast_fp16_to_fp32 = cast(dtype = var_465_cast_fp16_to_fp32_dtype_0, x = var_465_cast_fp16)[name = tensor<string, []>("cast_652")];
            tensor<fp32, [1, 1, 1]> var_466 = rsqrt(epsilon = var_466_epsilon_0, x = var_465_cast_fp16_to_fp32)[name = tensor<string, []>("op_466")];
            tensor<string, []> var_466_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_466_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_466_to_fp16 = cast(dtype = var_466_to_fp16_dtype_0, x = var_466)[name = tensor<string, []>("cast_651")];
            tensor<fp16, [1, 1, 512]> x_75_cast_fp16 = mul(x = x_69_cast_fp16, y = var_466_to_fp16)[name = tensor<string, []>("x_75_cast_fp16")];
            tensor<fp16, [512]> layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235570880)))];
            tensor<fp16, [1, 1, 512]> x_77_cast_fp16 = mul(x = layers_1_post_attention_layernorm_weight_to_fp16, y = x_75_cast_fp16)[name = tensor<string, []>("x_77_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235571968)))];
            tensor<fp16, [1, 1, 1536]> linear_11_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_mlp_gate_proj_weight_to_fp16, x = x_77_cast_fp16)[name = tensor<string, []>("linear_11_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_477_cast_fp16 = silu(x = linear_11_cast_fp16)[name = tensor<string, []>("op_477_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_1_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237144896)))];
            tensor<fp16, [1, 1, 1536]> linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_mlp_up_proj_weight_to_fp16, x = x_77_cast_fp16)[name = tensor<string, []>("linear_12_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_79_cast_fp16 = mul(x = var_477_cast_fp16, y = linear_12_cast_fp16)[name = tensor<string, []>("x_79_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238717824)))];
            tensor<fp16, [1, 1, 512]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_1_mlp_down_proj_weight_to_fp16, x = x_79_cast_fp16)[name = tensor<string, []>("linear_13_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_81_cast_fp16 = add(x = x_69_cast_fp16, y = linear_13_cast_fp16)[name = tensor<string, []>("x_81_cast_fp16")];
            tensor<string, []> x_81_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_81_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_5_begin_0 = const()[name = tensor<string, []>("k_cache_5_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_5_end_0 = const()[name = tensor<string, []>("k_cache_5_end_0"), val = tensor<int32, [5]>([3, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_5_end_mask_0 = const()[name = tensor<string, []>("k_cache_5_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_5_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_5_cast_fp16 = slice_by_index(begin = k_cache_5_begin_0, end = k_cache_5_end_0, end_mask = k_cache_5_end_mask_0, squeeze_mask = k_cache_5_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_5_cast_fp16")];
            tensor<int32, [5]> v_cache_5_begin_0 = const()[name = tensor<string, []>("v_cache_5_begin_0"), val = tensor<int32, [5]>([2, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_5_end_0 = const()[name = tensor<string, []>("v_cache_5_end_0"), val = tensor<int32, [5]>([3, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_5_end_mask_0 = const()[name = tensor<string, []>("v_cache_5_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_5_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_5_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_5_cast_fp16 = slice_by_index(begin = v_cache_5_begin_0, end = v_cache_5_end_0, end_mask = v_cache_5_end_mask_0, squeeze_mask = v_cache_5_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_5_cast_fp16")];
            tensor<int32, []> var_509 = const()[name = tensor<string, []>("op_509"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_508_promoted = const()[name = tensor<string, []>("op_508_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_81_cast_fp16_to_fp32 = cast(dtype = x_81_cast_fp16_to_fp32_dtype_0, x = x_81_cast_fp16)[name = tensor<string, []>("cast_650")];
            tensor<fp32, [1, 1, 512]> var_518 = pow(x = x_81_cast_fp16_to_fp32, y = var_508_promoted)[name = tensor<string, []>("op_518")];
            tensor<int32, [1]> var_17_axes_0 = const()[name = tensor<string, []>("var_17_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_17_keep_dims_0 = const()[name = tensor<string, []>("var_17_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_17 = reduce_mean(axes = var_17_axes_0, keep_dims = var_17_keep_dims_0, x = var_518)[name = tensor<string, []>("var_17")];
            tensor<string, []> var_17_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_17_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_522_to_fp16 = const()[name = tensor<string, []>("op_522_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_17_to_fp16 = cast(dtype = var_17_to_fp16_dtype_0, x = var_17)[name = tensor<string, []>("cast_649")];
            tensor<fp16, [1, 1, 1]> var_523_cast_fp16 = add(x = var_17_to_fp16, y = var_522_to_fp16)[name = tensor<string, []>("op_523_cast_fp16")];
            tensor<string, []> var_523_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_523_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_524_epsilon_0 = const()[name = tensor<string, []>("op_524_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_523_cast_fp16_to_fp32 = cast(dtype = var_523_cast_fp16_to_fp32_dtype_0, x = var_523_cast_fp16)[name = tensor<string, []>("cast_648")];
            tensor<fp32, [1, 1, 1]> var_524 = rsqrt(epsilon = var_524_epsilon_0, x = var_523_cast_fp16_to_fp32)[name = tensor<string, []>("op_524")];
            tensor<string, []> var_524_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_524_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_524_to_fp16 = cast(dtype = var_524_to_fp16_dtype_0, x = var_524)[name = tensor<string, []>("cast_647")];
            tensor<fp16, [1, 1, 512]> x_87_cast_fp16 = mul(x = x_81_cast_fp16, y = var_524_to_fp16)[name = tensor<string, []>("x_87_cast_fp16")];
            tensor<fp16, [512]> layers_2_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240290752)))];
            tensor<fp16, [1, 1, 512]> x_89_cast_fp16 = mul(x = layers_2_input_layernorm_weight_to_fp16, y = x_87_cast_fp16)[name = tensor<string, []>("x_89_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(240291840)))];
            tensor<fp16, [1, 1, 1536]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = x_89_cast_fp16)[name = tensor<string, []>("linear_14_cast_fp16")];
            tensor<int32, [4]> var_538 = const()[name = tensor<string, []>("op_538"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_91_cast_fp16 = reshape(shape = var_538, x = linear_14_cast_fp16)[name = tensor<string, []>("x_91_cast_fp16")];
            tensor<string, []> x_91_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_91_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_508_promoted_1 = const()[name = tensor<string, []>("op_508_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_91_cast_fp16_to_fp32 = cast(dtype = x_91_cast_fp16_to_fp32_dtype_0, x = x_91_cast_fp16)[name = tensor<string, []>("cast_646")];
            tensor<fp32, [1, 1, 12, 128]> var_542 = pow(x = x_91_cast_fp16_to_fp32, y = var_508_promoted_1)[name = tensor<string, []>("op_542")];
            tensor<int32, [1]> var_19_axes_0 = const()[name = tensor<string, []>("var_19_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_19_keep_dims_0 = const()[name = tensor<string, []>("var_19_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_19 = reduce_mean(axes = var_19_axes_0, keep_dims = var_19_keep_dims_0, x = var_542)[name = tensor<string, []>("var_19")];
            tensor<string, []> var_19_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_19_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_546_to_fp16 = const()[name = tensor<string, []>("op_546_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_19_to_fp16 = cast(dtype = var_19_to_fp16_dtype_0, x = var_19)[name = tensor<string, []>("cast_645")];
            tensor<fp16, [1, 1, 12, 1]> var_547_cast_fp16 = add(x = var_19_to_fp16, y = var_546_to_fp16)[name = tensor<string, []>("op_547_cast_fp16")];
            tensor<string, []> var_547_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_547_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_548_epsilon_0 = const()[name = tensor<string, []>("op_548_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_547_cast_fp16_to_fp32 = cast(dtype = var_547_cast_fp16_to_fp32_dtype_0, x = var_547_cast_fp16)[name = tensor<string, []>("cast_644")];
            tensor<fp32, [1, 1, 12, 1]> var_548 = rsqrt(epsilon = var_548_epsilon_0, x = var_547_cast_fp16_to_fp32)[name = tensor<string, []>("op_548")];
            tensor<string, []> var_548_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_548_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_548_to_fp16 = cast(dtype = var_548_to_fp16_dtype_0, x = var_548)[name = tensor<string, []>("cast_643")];
            tensor<fp16, [1, 1, 12, 128]> x_97_cast_fp16 = mul(x = x_91_cast_fp16, y = var_548_to_fp16)[name = tensor<string, []>("x_97_cast_fp16")];
            tensor<fp16, [128]> layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241864768)))];
            tensor<fp16, [1, 1, 12, 128]> var_550_cast_fp16 = mul(x = layers_2_self_attn_q_norm_weight_to_fp16, y = x_97_cast_fp16)[name = tensor<string, []>("op_550_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = tensor<string, []>("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(241865088)))];
            tensor<fp16, [1, 1, 512]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = x_89_cast_fp16)[name = tensor<string, []>("linear_15_cast_fp16")];
            tensor<int32, [4]> var_554 = const()[name = tensor<string, []>("op_554"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_99_cast_fp16 = reshape(shape = var_554, x = linear_15_cast_fp16)[name = tensor<string, []>("x_99_cast_fp16")];
            tensor<string, []> x_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_99_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_508_promoted_2 = const()[name = tensor<string, []>("op_508_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_99_cast_fp16_to_fp32 = cast(dtype = x_99_cast_fp16_to_fp32_dtype_0, x = x_99_cast_fp16)[name = tensor<string, []>("cast_642")];
            tensor<fp32, [1, 1, 4, 128]> var_558 = pow(x = x_99_cast_fp16_to_fp32, y = var_508_promoted_2)[name = tensor<string, []>("op_558")];
            tensor<int32, [1]> var_21_axes_0 = const()[name = tensor<string, []>("var_21_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_21_keep_dims_0 = const()[name = tensor<string, []>("var_21_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_21 = reduce_mean(axes = var_21_axes_0, keep_dims = var_21_keep_dims_0, x = var_558)[name = tensor<string, []>("var_21")];
            tensor<string, []> var_21_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_21_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_562_to_fp16 = const()[name = tensor<string, []>("op_562_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_21_to_fp16 = cast(dtype = var_21_to_fp16_dtype_0, x = var_21)[name = tensor<string, []>("cast_641")];
            tensor<fp16, [1, 1, 4, 1]> var_563_cast_fp16 = add(x = var_21_to_fp16, y = var_562_to_fp16)[name = tensor<string, []>("op_563_cast_fp16")];
            tensor<string, []> var_563_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_563_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_564_epsilon_0 = const()[name = tensor<string, []>("op_564_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_563_cast_fp16_to_fp32 = cast(dtype = var_563_cast_fp16_to_fp32_dtype_0, x = var_563_cast_fp16)[name = tensor<string, []>("cast_640")];
            tensor<fp32, [1, 1, 4, 1]> var_564 = rsqrt(epsilon = var_564_epsilon_0, x = var_563_cast_fp16_to_fp32)[name = tensor<string, []>("op_564")];
            tensor<string, []> var_564_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_564_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_564_to_fp16 = cast(dtype = var_564_to_fp16_dtype_0, x = var_564)[name = tensor<string, []>("cast_639")];
            tensor<fp16, [1, 1, 4, 128]> x_105_cast_fp16 = mul(x = x_99_cast_fp16, y = var_564_to_fp16)[name = tensor<string, []>("x_105_cast_fp16")];
            tensor<fp16, [128]> layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242389440)))];
            tensor<fp16, [1, 1, 4, 128]> var_566_cast_fp16 = mul(x = layers_2_self_attn_k_norm_weight_to_fp16, y = x_105_cast_fp16)[name = tensor<string, []>("op_566_cast_fp16")];
            tensor<int32, [4]> k_9_perm_0 = const()[name = tensor<string, []>("k_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242389760)))];
            tensor<fp16, [1, 1, 512]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = x_89_cast_fp16)[name = tensor<string, []>("linear_16_cast_fp16")];
            tensor<int32, [4]> var_570 = const()[name = tensor<string, []>("op_570"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_571_cast_fp16 = reshape(shape = var_570, x = linear_16_cast_fp16)[name = tensor<string, []>("op_571_cast_fp16")];
            tensor<int32, [4]> v_5_perm_0 = const()[name = tensor<string, []>("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = var_550_cast_fp16)[name = tensor<string, []>("transpose_103")];
            tensor<fp16, [1, 12, 1, 128]> var_575_cast_fp16 = mul(x = q_9_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_575_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = tensor<string, []>("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = tensor<string, []>("x1_9_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = tensor<string, []>("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_9_cast_fp16)[name = tensor<string, []>("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = tensor<string, []>("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = tensor<string, []>("x2_9_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = tensor<string, []>("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_9_cast_fp16)[name = tensor<string, []>("x2_9_cast_fp16")];
            tensor<fp16, []> const_5_promoted_to_fp16 = const()[name = tensor<string, []>("const_5_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_578_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_5_promoted_to_fp16)[name = tensor<string, []>("op_578_cast_fp16")];
            tensor<bool, []> var_580_interleave_0 = const()[name = tensor<string, []>("op_580_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_580_cast_fp16 = concat(axis = var_509, interleave = var_580_interleave_0, values = (var_578_cast_fp16, x1_9_cast_fp16))[name = tensor<string, []>("op_580_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_581_cast_fp16 = mul(x = var_580_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_581_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_11_cast_fp16 = add(x = var_575_cast_fp16, y = var_581_cast_fp16)[name = tensor<string, []>("q_11_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_9_cast_fp16 = transpose(perm = k_9_perm_0, x = var_566_cast_fp16)[name = tensor<string, []>("transpose_102")];
            tensor<fp16, [1, 4, 1, 128]> var_583_cast_fp16 = mul(x = k_9_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_583_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = tensor<string, []>("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = tensor<string, []>("x1_11_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = tensor<string, []>("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_9_cast_fp16)[name = tensor<string, []>("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = tensor<string, []>("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = tensor<string, []>("x2_11_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = tensor<string, []>("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_9_cast_fp16)[name = tensor<string, []>("x2_11_cast_fp16")];
            tensor<fp16, []> const_6_promoted_to_fp16 = const()[name = tensor<string, []>("const_6_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_586_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_6_promoted_to_fp16)[name = tensor<string, []>("op_586_cast_fp16")];
            tensor<bool, []> var_588_interleave_0 = const()[name = tensor<string, []>("op_588_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_588_cast_fp16 = concat(axis = var_509, interleave = var_588_interleave_0, values = (var_586_cast_fp16, x1_11_cast_fp16))[name = tensor<string, []>("op_588_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_589_cast_fp16 = mul(x = var_588_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_589_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_11_cast_fp16 = add(x = var_583_cast_fp16, y = var_589_cast_fp16)[name = tensor<string, []>("k_11_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_592_cast_fp16 = mul(x = k_cache_5_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_592_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_593_cast_fp16 = mul(x = k_11_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_593_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_5_cast_fp16 = add(x = var_592_cast_fp16, y = var_593_cast_fp16)[name = tensor<string, []>("k_full_5_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_596_cast_fp16 = mul(x = v_cache_5_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_596_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_571_cast_fp16)[name = tensor<string, []>("transpose_101")];
            tensor<fp16, [1, 4, 2048, 128]> var_597_cast_fp16 = mul(x = v_5_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_597_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_5_cast_fp16 = add(x = var_596_cast_fp16, y = var_597_cast_fp16)[name = tensor<string, []>("v_full_5_cast_fp16")];
            tensor<int32, [1]> var_599_axes_0 = const()[name = tensor<string, []>("op_599_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_599_cast_fp16 = expand_dims(axes = var_599_axes_0, x = k_full_5_cast_fp16)[name = tensor<string, []>("op_599_cast_fp16")];
            tensor<int32, [5]> var_601_reps_0 = const()[name = tensor<string, []>("op_601_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_601_cast_fp16 = tile(reps = var_601_reps_0, x = var_599_cast_fp16)[name = tensor<string, []>("op_601_cast_fp16")];
            tensor<int32, [4]> var_602 = const()[name = tensor<string, []>("op_602"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_5_cast_fp16 = reshape(shape = var_602, x = var_601_cast_fp16)[name = tensor<string, []>("k_rep_5_cast_fp16")];
            tensor<int32, [1]> var_604_axes_0 = const()[name = tensor<string, []>("op_604_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_604_cast_fp16 = expand_dims(axes = var_604_axes_0, x = v_full_5_cast_fp16)[name = tensor<string, []>("op_604_cast_fp16")];
            tensor<int32, [5]> var_606_reps_0 = const()[name = tensor<string, []>("op_606_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_606_cast_fp16 = tile(reps = var_606_reps_0, x = var_604_cast_fp16)[name = tensor<string, []>("op_606_cast_fp16")];
            tensor<int32, [4]> var_607 = const()[name = tensor<string, []>("op_607"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_5_cast_fp16 = reshape(shape = var_607, x = var_606_cast_fp16)[name = tensor<string, []>("v_rep_5_cast_fp16")];
            tensor<bool, []> var_610_transpose_x_1 = const()[name = tensor<string, []>("op_610_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_610_transpose_y_1 = const()[name = tensor<string, []>("op_610_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_610_cast_fp16 = matmul(transpose_x = var_610_transpose_x_1, transpose_y = var_610_transpose_y_1, x = q_11_cast_fp16, y = k_rep_5_cast_fp16)[name = tensor<string, []>("op_610_cast_fp16")];
            tensor<fp16, []> var_611_to_fp16 = const()[name = tensor<string, []>("op_611_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_9_cast_fp16 = mul(x = var_610_cast_fp16, y = var_611_to_fp16)[name = tensor<string, []>("attn_9_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_11_cast_fp16 = add(x = attn_9_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
            tensor<string, []> input_11_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_11_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_11_cast_fp16_to_fp32 = cast(dtype = input_11_cast_fp16_to_fp32_dtype_0, x = input_11_cast_fp16)[name = tensor<string, []>("cast_638")];
            tensor<fp32, [1, 12, 1, 2048]> attn_11 = softmax(axis = var_509, x = input_11_cast_fp16_to_fp32)[name = tensor<string, []>("attn_11")];
            tensor<bool, []> out_5_transpose_x_0 = const()[name = tensor<string, []>("out_5_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_5_transpose_y_0 = const()[name = tensor<string, []>("out_5_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_11_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_11_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_11_to_fp16 = cast(dtype = attn_11_to_fp16_dtype_0, x = attn_11)[name = tensor<string, []>("cast_637")];
            tensor<fp16, [1, 12, 1, 128]> out_5_cast_fp16 = matmul(transpose_x = out_5_transpose_x_0, transpose_y = out_5_transpose_y_0, x = attn_11_to_fp16, y = v_rep_5_cast_fp16)[name = tensor<string, []>("out_5_cast_fp16")];
            tensor<int32, [4]> var_616_perm_0 = const()[name = tensor<string, []>("op_616_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_618 = const()[name = tensor<string, []>("op_618"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_616_cast_fp16 = transpose(perm = var_616_perm_0, x = out_5_cast_fp16)[name = tensor<string, []>("transpose_100")];
            tensor<fp16, [1, 1, 1536]> x_107_cast_fp16 = reshape(shape = var_618, x = var_616_cast_fp16)[name = tensor<string, []>("x_107_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(242914112)))];
            tensor<fp16, [1, 1, 512]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = x_107_cast_fp16)[name = tensor<string, []>("linear_17_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_109_cast_fp16 = add(x = x_81_cast_fp16, y = linear_17_cast_fp16)[name = tensor<string, []>("x_109_cast_fp16")];
            tensor<string, []> x_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_109_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_508_promoted_3 = const()[name = tensor<string, []>("op_508_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_109_cast_fp16_to_fp32 = cast(dtype = x_109_cast_fp16_to_fp32_dtype_0, x = x_109_cast_fp16)[name = tensor<string, []>("cast_636")];
            tensor<fp32, [1, 1, 512]> var_629 = pow(x = x_109_cast_fp16_to_fp32, y = var_508_promoted_3)[name = tensor<string, []>("op_629")];
            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("var_23_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_23_keep_dims_0 = const()[name = tensor<string, []>("var_23_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_23 = reduce_mean(axes = var_23_axes_0, keep_dims = var_23_keep_dims_0, x = var_629)[name = tensor<string, []>("var_23")];
            tensor<string, []> var_23_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_23_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_633_to_fp16 = const()[name = tensor<string, []>("op_633_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_23_to_fp16 = cast(dtype = var_23_to_fp16_dtype_0, x = var_23)[name = tensor<string, []>("cast_635")];
            tensor<fp16, [1, 1, 1]> var_634_cast_fp16 = add(x = var_23_to_fp16, y = var_633_to_fp16)[name = tensor<string, []>("op_634_cast_fp16")];
            tensor<string, []> var_634_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_634_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_635_epsilon_0 = const()[name = tensor<string, []>("op_635_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_634_cast_fp16_to_fp32 = cast(dtype = var_634_cast_fp16_to_fp32_dtype_0, x = var_634_cast_fp16)[name = tensor<string, []>("cast_634")];
            tensor<fp32, [1, 1, 1]> var_635 = rsqrt(epsilon = var_635_epsilon_0, x = var_634_cast_fp16_to_fp32)[name = tensor<string, []>("op_635")];
            tensor<string, []> var_635_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_635_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_635_to_fp16 = cast(dtype = var_635_to_fp16_dtype_0, x = var_635)[name = tensor<string, []>("cast_633")];
            tensor<fp16, [1, 1, 512]> x_115_cast_fp16 = mul(x = x_109_cast_fp16, y = var_635_to_fp16)[name = tensor<string, []>("x_115_cast_fp16")];
            tensor<fp16, [512]> layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244487040)))];
            tensor<fp16, [1, 1, 512]> x_117_cast_fp16 = mul(x = layers_2_post_attention_layernorm_weight_to_fp16, y = x_115_cast_fp16)[name = tensor<string, []>("x_117_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(244488128)))];
            tensor<fp16, [1, 1, 1536]> linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_mlp_gate_proj_weight_to_fp16, x = x_117_cast_fp16)[name = tensor<string, []>("linear_18_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_646_cast_fp16 = silu(x = linear_18_cast_fp16)[name = tensor<string, []>("op_646_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_2_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(246061056)))];
            tensor<fp16, [1, 1, 1536]> linear_19_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_mlp_up_proj_weight_to_fp16, x = x_117_cast_fp16)[name = tensor<string, []>("linear_19_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_119_cast_fp16 = mul(x = var_646_cast_fp16, y = linear_19_cast_fp16)[name = tensor<string, []>("x_119_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247633984)))];
            tensor<fp16, [1, 1, 512]> linear_20_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_2_mlp_down_proj_weight_to_fp16, x = x_119_cast_fp16)[name = tensor<string, []>("linear_20_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_121_cast_fp16 = add(x = x_109_cast_fp16, y = linear_20_cast_fp16)[name = tensor<string, []>("x_121_cast_fp16")];
            tensor<string, []> x_121_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_121_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_7_begin_0 = const()[name = tensor<string, []>("k_cache_7_begin_0"), val = tensor<int32, [5]>([3, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_7_end_0 = const()[name = tensor<string, []>("k_cache_7_end_0"), val = tensor<int32, [5]>([4, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_7_end_mask_0 = const()[name = tensor<string, []>("k_cache_7_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_7_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_7_cast_fp16 = slice_by_index(begin = k_cache_7_begin_0, end = k_cache_7_end_0, end_mask = k_cache_7_end_mask_0, squeeze_mask = k_cache_7_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_7_cast_fp16")];
            tensor<int32, [5]> v_cache_7_begin_0 = const()[name = tensor<string, []>("v_cache_7_begin_0"), val = tensor<int32, [5]>([3, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_7_end_0 = const()[name = tensor<string, []>("v_cache_7_end_0"), val = tensor<int32, [5]>([4, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_7_end_mask_0 = const()[name = tensor<string, []>("v_cache_7_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_7_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_7_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_7_cast_fp16 = slice_by_index(begin = v_cache_7_begin_0, end = v_cache_7_end_0, end_mask = v_cache_7_end_mask_0, squeeze_mask = v_cache_7_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_7_cast_fp16")];
            tensor<int32, []> var_678 = const()[name = tensor<string, []>("op_678"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_677_promoted = const()[name = tensor<string, []>("op_677_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_121_cast_fp16_to_fp32 = cast(dtype = x_121_cast_fp16_to_fp32_dtype_0, x = x_121_cast_fp16)[name = tensor<string, []>("cast_632")];
            tensor<fp32, [1, 1, 512]> var_687 = pow(x = x_121_cast_fp16_to_fp32, y = var_677_promoted)[name = tensor<string, []>("op_687")];
            tensor<int32, [1]> var_25_axes_0 = const()[name = tensor<string, []>("var_25_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_25_keep_dims_0 = const()[name = tensor<string, []>("var_25_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_25 = reduce_mean(axes = var_25_axes_0, keep_dims = var_25_keep_dims_0, x = var_687)[name = tensor<string, []>("var_25")];
            tensor<string, []> var_25_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_25_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_691_to_fp16 = const()[name = tensor<string, []>("op_691_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_25_to_fp16 = cast(dtype = var_25_to_fp16_dtype_0, x = var_25)[name = tensor<string, []>("cast_631")];
            tensor<fp16, [1, 1, 1]> var_692_cast_fp16 = add(x = var_25_to_fp16, y = var_691_to_fp16)[name = tensor<string, []>("op_692_cast_fp16")];
            tensor<string, []> var_692_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_692_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_693_epsilon_0 = const()[name = tensor<string, []>("op_693_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_692_cast_fp16_to_fp32 = cast(dtype = var_692_cast_fp16_to_fp32_dtype_0, x = var_692_cast_fp16)[name = tensor<string, []>("cast_630")];
            tensor<fp32, [1, 1, 1]> var_693 = rsqrt(epsilon = var_693_epsilon_0, x = var_692_cast_fp16_to_fp32)[name = tensor<string, []>("op_693")];
            tensor<string, []> var_693_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_693_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_693_to_fp16 = cast(dtype = var_693_to_fp16_dtype_0, x = var_693)[name = tensor<string, []>("cast_629")];
            tensor<fp16, [1, 1, 512]> x_127_cast_fp16 = mul(x = x_121_cast_fp16, y = var_693_to_fp16)[name = tensor<string, []>("x_127_cast_fp16")];
            tensor<fp16, [512]> layers_3_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249206912)))];
            tensor<fp16, [1, 1, 512]> x_129_cast_fp16 = mul(x = layers_3_input_layernorm_weight_to_fp16, y = x_127_cast_fp16)[name = tensor<string, []>("x_129_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(249208000)))];
            tensor<fp16, [1, 1, 1536]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = x_129_cast_fp16)[name = tensor<string, []>("linear_21_cast_fp16")];
            tensor<int32, [4]> var_707 = const()[name = tensor<string, []>("op_707"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_131_cast_fp16 = reshape(shape = var_707, x = linear_21_cast_fp16)[name = tensor<string, []>("x_131_cast_fp16")];
            tensor<string, []> x_131_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_131_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_677_promoted_1 = const()[name = tensor<string, []>("op_677_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_131_cast_fp16_to_fp32 = cast(dtype = x_131_cast_fp16_to_fp32_dtype_0, x = x_131_cast_fp16)[name = tensor<string, []>("cast_628")];
            tensor<fp32, [1, 1, 12, 128]> var_711 = pow(x = x_131_cast_fp16_to_fp32, y = var_677_promoted_1)[name = tensor<string, []>("op_711")];
            tensor<int32, [1]> var_27_axes_0 = const()[name = tensor<string, []>("var_27_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_27_keep_dims_0 = const()[name = tensor<string, []>("var_27_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_27 = reduce_mean(axes = var_27_axes_0, keep_dims = var_27_keep_dims_0, x = var_711)[name = tensor<string, []>("var_27")];
            tensor<string, []> var_27_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_27_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_715_to_fp16 = const()[name = tensor<string, []>("op_715_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_27_to_fp16 = cast(dtype = var_27_to_fp16_dtype_0, x = var_27)[name = tensor<string, []>("cast_627")];
            tensor<fp16, [1, 1, 12, 1]> var_716_cast_fp16 = add(x = var_27_to_fp16, y = var_715_to_fp16)[name = tensor<string, []>("op_716_cast_fp16")];
            tensor<string, []> var_716_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_716_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_717_epsilon_0 = const()[name = tensor<string, []>("op_717_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_716_cast_fp16_to_fp32 = cast(dtype = var_716_cast_fp16_to_fp32_dtype_0, x = var_716_cast_fp16)[name = tensor<string, []>("cast_626")];
            tensor<fp32, [1, 1, 12, 1]> var_717 = rsqrt(epsilon = var_717_epsilon_0, x = var_716_cast_fp16_to_fp32)[name = tensor<string, []>("op_717")];
            tensor<string, []> var_717_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_717_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_717_to_fp16 = cast(dtype = var_717_to_fp16_dtype_0, x = var_717)[name = tensor<string, []>("cast_625")];
            tensor<fp16, [1, 1, 12, 128]> x_137_cast_fp16 = mul(x = x_131_cast_fp16, y = var_717_to_fp16)[name = tensor<string, []>("x_137_cast_fp16")];
            tensor<fp16, [128]> layers_3_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250780928)))];
            tensor<fp16, [1, 1, 12, 128]> var_719_cast_fp16 = mul(x = layers_3_self_attn_q_norm_weight_to_fp16, y = x_137_cast_fp16)[name = tensor<string, []>("op_719_cast_fp16")];
            tensor<int32, [4]> q_13_perm_0 = const()[name = tensor<string, []>("q_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(250781248)))];
            tensor<fp16, [1, 1, 512]> linear_22_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = x_129_cast_fp16)[name = tensor<string, []>("linear_22_cast_fp16")];
            tensor<int32, [4]> var_723 = const()[name = tensor<string, []>("op_723"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_139_cast_fp16 = reshape(shape = var_723, x = linear_22_cast_fp16)[name = tensor<string, []>("x_139_cast_fp16")];
            tensor<string, []> x_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_139_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_677_promoted_2 = const()[name = tensor<string, []>("op_677_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_139_cast_fp16_to_fp32 = cast(dtype = x_139_cast_fp16_to_fp32_dtype_0, x = x_139_cast_fp16)[name = tensor<string, []>("cast_624")];
            tensor<fp32, [1, 1, 4, 128]> var_727 = pow(x = x_139_cast_fp16_to_fp32, y = var_677_promoted_2)[name = tensor<string, []>("op_727")];
            tensor<int32, [1]> var_29_axes_0 = const()[name = tensor<string, []>("var_29_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_29_keep_dims_0 = const()[name = tensor<string, []>("var_29_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_29 = reduce_mean(axes = var_29_axes_0, keep_dims = var_29_keep_dims_0, x = var_727)[name = tensor<string, []>("var_29")];
            tensor<string, []> var_29_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_29_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_731_to_fp16 = const()[name = tensor<string, []>("op_731_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_29_to_fp16 = cast(dtype = var_29_to_fp16_dtype_0, x = var_29)[name = tensor<string, []>("cast_623")];
            tensor<fp16, [1, 1, 4, 1]> var_732_cast_fp16 = add(x = var_29_to_fp16, y = var_731_to_fp16)[name = tensor<string, []>("op_732_cast_fp16")];
            tensor<string, []> var_732_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_732_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_733_epsilon_0 = const()[name = tensor<string, []>("op_733_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_732_cast_fp16_to_fp32 = cast(dtype = var_732_cast_fp16_to_fp32_dtype_0, x = var_732_cast_fp16)[name = tensor<string, []>("cast_622")];
            tensor<fp32, [1, 1, 4, 1]> var_733 = rsqrt(epsilon = var_733_epsilon_0, x = var_732_cast_fp16_to_fp32)[name = tensor<string, []>("op_733")];
            tensor<string, []> var_733_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_733_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_733_to_fp16 = cast(dtype = var_733_to_fp16_dtype_0, x = var_733)[name = tensor<string, []>("cast_621")];
            tensor<fp16, [1, 1, 4, 128]> x_145_cast_fp16 = mul(x = x_139_cast_fp16, y = var_733_to_fp16)[name = tensor<string, []>("x_145_cast_fp16")];
            tensor<fp16, [128]> layers_3_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251305600)))];
            tensor<fp16, [1, 1, 4, 128]> var_735_cast_fp16 = mul(x = layers_3_self_attn_k_norm_weight_to_fp16, y = x_145_cast_fp16)[name = tensor<string, []>("op_735_cast_fp16")];
            tensor<int32, [4]> k_13_perm_0 = const()[name = tensor<string, []>("k_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251305920)))];
            tensor<fp16, [1, 1, 512]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = x_129_cast_fp16)[name = tensor<string, []>("linear_23_cast_fp16")];
            tensor<int32, [4]> var_739 = const()[name = tensor<string, []>("op_739"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_740_cast_fp16 = reshape(shape = var_739, x = linear_23_cast_fp16)[name = tensor<string, []>("op_740_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = tensor<string, []>("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_13_cast_fp16 = transpose(perm = q_13_perm_0, x = var_719_cast_fp16)[name = tensor<string, []>("transpose_99")];
            tensor<fp16, [1, 12, 1, 128]> var_744_cast_fp16 = mul(x = q_13_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_744_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = tensor<string, []>("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = tensor<string, []>("x1_13_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = tensor<string, []>("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_13_cast_fp16)[name = tensor<string, []>("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = tensor<string, []>("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = tensor<string, []>("x2_13_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = tensor<string, []>("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_13_cast_fp16)[name = tensor<string, []>("x2_13_cast_fp16")];
            tensor<fp16, []> const_7_promoted_to_fp16 = const()[name = tensor<string, []>("const_7_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_747_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_7_promoted_to_fp16)[name = tensor<string, []>("op_747_cast_fp16")];
            tensor<bool, []> var_749_interleave_0 = const()[name = tensor<string, []>("op_749_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_749_cast_fp16 = concat(axis = var_678, interleave = var_749_interleave_0, values = (var_747_cast_fp16, x1_13_cast_fp16))[name = tensor<string, []>("op_749_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_750_cast_fp16 = mul(x = var_749_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_750_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_15_cast_fp16 = add(x = var_744_cast_fp16, y = var_750_cast_fp16)[name = tensor<string, []>("q_15_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_13_cast_fp16 = transpose(perm = k_13_perm_0, x = var_735_cast_fp16)[name = tensor<string, []>("transpose_98")];
            tensor<fp16, [1, 4, 1, 128]> var_752_cast_fp16 = mul(x = k_13_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_752_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = tensor<string, []>("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = tensor<string, []>("x1_15_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = tensor<string, []>("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_13_cast_fp16)[name = tensor<string, []>("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = tensor<string, []>("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = tensor<string, []>("x2_15_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = tensor<string, []>("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_13_cast_fp16)[name = tensor<string, []>("x2_15_cast_fp16")];
            tensor<fp16, []> const_8_promoted_to_fp16 = const()[name = tensor<string, []>("const_8_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_755_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_8_promoted_to_fp16)[name = tensor<string, []>("op_755_cast_fp16")];
            tensor<bool, []> var_757_interleave_0 = const()[name = tensor<string, []>("op_757_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_757_cast_fp16 = concat(axis = var_678, interleave = var_757_interleave_0, values = (var_755_cast_fp16, x1_15_cast_fp16))[name = tensor<string, []>("op_757_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_758_cast_fp16 = mul(x = var_757_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_758_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_15_cast_fp16 = add(x = var_752_cast_fp16, y = var_758_cast_fp16)[name = tensor<string, []>("k_15_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_761_cast_fp16 = mul(x = k_cache_7_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_761_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_762_cast_fp16 = mul(x = k_15_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_762_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_7_cast_fp16 = add(x = var_761_cast_fp16, y = var_762_cast_fp16)[name = tensor<string, []>("k_full_7_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_765_cast_fp16 = mul(x = v_cache_7_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_765_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_740_cast_fp16)[name = tensor<string, []>("transpose_97")];
            tensor<fp16, [1, 4, 2048, 128]> var_766_cast_fp16 = mul(x = v_7_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_766_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_7_cast_fp16 = add(x = var_765_cast_fp16, y = var_766_cast_fp16)[name = tensor<string, []>("v_full_7_cast_fp16")];
            tensor<int32, [1]> var_768_axes_0 = const()[name = tensor<string, []>("op_768_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_768_cast_fp16 = expand_dims(axes = var_768_axes_0, x = k_full_7_cast_fp16)[name = tensor<string, []>("op_768_cast_fp16")];
            tensor<int32, [5]> var_770_reps_0 = const()[name = tensor<string, []>("op_770_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_770_cast_fp16 = tile(reps = var_770_reps_0, x = var_768_cast_fp16)[name = tensor<string, []>("op_770_cast_fp16")];
            tensor<int32, [4]> var_771 = const()[name = tensor<string, []>("op_771"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_7_cast_fp16 = reshape(shape = var_771, x = var_770_cast_fp16)[name = tensor<string, []>("k_rep_7_cast_fp16")];
            tensor<int32, [1]> var_773_axes_0 = const()[name = tensor<string, []>("op_773_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_773_cast_fp16 = expand_dims(axes = var_773_axes_0, x = v_full_7_cast_fp16)[name = tensor<string, []>("op_773_cast_fp16")];
            tensor<int32, [5]> var_775_reps_0 = const()[name = tensor<string, []>("op_775_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_775_cast_fp16 = tile(reps = var_775_reps_0, x = var_773_cast_fp16)[name = tensor<string, []>("op_775_cast_fp16")];
            tensor<int32, [4]> var_776 = const()[name = tensor<string, []>("op_776"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_7_cast_fp16 = reshape(shape = var_776, x = var_775_cast_fp16)[name = tensor<string, []>("v_rep_7_cast_fp16")];
            tensor<bool, []> var_779_transpose_x_1 = const()[name = tensor<string, []>("op_779_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_779_transpose_y_1 = const()[name = tensor<string, []>("op_779_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_779_cast_fp16 = matmul(transpose_x = var_779_transpose_x_1, transpose_y = var_779_transpose_y_1, x = q_15_cast_fp16, y = k_rep_7_cast_fp16)[name = tensor<string, []>("op_779_cast_fp16")];
            tensor<fp16, []> var_780_to_fp16 = const()[name = tensor<string, []>("op_780_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_13_cast_fp16 = mul(x = var_779_cast_fp16, y = var_780_to_fp16)[name = tensor<string, []>("attn_13_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_15_cast_fp16 = add(x = attn_13_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
            tensor<string, []> input_15_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_15_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_15_cast_fp16_to_fp32 = cast(dtype = input_15_cast_fp16_to_fp32_dtype_0, x = input_15_cast_fp16)[name = tensor<string, []>("cast_620")];
            tensor<fp32, [1, 12, 1, 2048]> attn_15 = softmax(axis = var_678, x = input_15_cast_fp16_to_fp32)[name = tensor<string, []>("attn_15")];
            tensor<bool, []> out_7_transpose_x_0 = const()[name = tensor<string, []>("out_7_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_7_transpose_y_0 = const()[name = tensor<string, []>("out_7_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_15_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_15_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_15_to_fp16 = cast(dtype = attn_15_to_fp16_dtype_0, x = attn_15)[name = tensor<string, []>("cast_619")];
            tensor<fp16, [1, 12, 1, 128]> out_7_cast_fp16 = matmul(transpose_x = out_7_transpose_x_0, transpose_y = out_7_transpose_y_0, x = attn_15_to_fp16, y = v_rep_7_cast_fp16)[name = tensor<string, []>("out_7_cast_fp16")];
            tensor<int32, [4]> var_785_perm_0 = const()[name = tensor<string, []>("op_785_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_787 = const()[name = tensor<string, []>("op_787"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_785_cast_fp16 = transpose(perm = var_785_perm_0, x = out_7_cast_fp16)[name = tensor<string, []>("transpose_96")];
            tensor<fp16, [1, 1, 1536]> x_147_cast_fp16 = reshape(shape = var_787, x = var_785_cast_fp16)[name = tensor<string, []>("x_147_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251830272)))];
            tensor<fp16, [1, 1, 512]> linear_24_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = x_147_cast_fp16)[name = tensor<string, []>("linear_24_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_149_cast_fp16 = add(x = x_121_cast_fp16, y = linear_24_cast_fp16)[name = tensor<string, []>("x_149_cast_fp16")];
            tensor<string, []> x_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_149_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_677_promoted_3 = const()[name = tensor<string, []>("op_677_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_149_cast_fp16_to_fp32 = cast(dtype = x_149_cast_fp16_to_fp32_dtype_0, x = x_149_cast_fp16)[name = tensor<string, []>("cast_618")];
            tensor<fp32, [1, 1, 512]> var_798 = pow(x = x_149_cast_fp16_to_fp32, y = var_677_promoted_3)[name = tensor<string, []>("op_798")];
            tensor<int32, [1]> var_31_axes_0 = const()[name = tensor<string, []>("var_31_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_31_keep_dims_0 = const()[name = tensor<string, []>("var_31_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_31 = reduce_mean(axes = var_31_axes_0, keep_dims = var_31_keep_dims_0, x = var_798)[name = tensor<string, []>("var_31")];
            tensor<string, []> var_31_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_31_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_802_to_fp16 = const()[name = tensor<string, []>("op_802_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_31_to_fp16 = cast(dtype = var_31_to_fp16_dtype_0, x = var_31)[name = tensor<string, []>("cast_617")];
            tensor<fp16, [1, 1, 1]> var_803_cast_fp16 = add(x = var_31_to_fp16, y = var_802_to_fp16)[name = tensor<string, []>("op_803_cast_fp16")];
            tensor<string, []> var_803_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_803_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_804_epsilon_0 = const()[name = tensor<string, []>("op_804_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_803_cast_fp16_to_fp32 = cast(dtype = var_803_cast_fp16_to_fp32_dtype_0, x = var_803_cast_fp16)[name = tensor<string, []>("cast_616")];
            tensor<fp32, [1, 1, 1]> var_804 = rsqrt(epsilon = var_804_epsilon_0, x = var_803_cast_fp16_to_fp32)[name = tensor<string, []>("op_804")];
            tensor<string, []> var_804_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_804_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_804_to_fp16 = cast(dtype = var_804_to_fp16_dtype_0, x = var_804)[name = tensor<string, []>("cast_615")];
            tensor<fp16, [1, 1, 512]> x_155_cast_fp16 = mul(x = x_149_cast_fp16, y = var_804_to_fp16)[name = tensor<string, []>("x_155_cast_fp16")];
            tensor<fp16, [512]> layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253403200)))];
            tensor<fp16, [1, 1, 512]> x_157_cast_fp16 = mul(x = layers_3_post_attention_layernorm_weight_to_fp16, y = x_155_cast_fp16)[name = tensor<string, []>("x_157_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(253404288)))];
            tensor<fp16, [1, 1, 1536]> linear_25_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_mlp_gate_proj_weight_to_fp16, x = x_157_cast_fp16)[name = tensor<string, []>("linear_25_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_815_cast_fp16 = silu(x = linear_25_cast_fp16)[name = tensor<string, []>("op_815_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(254977216)))];
            tensor<fp16, [1, 1, 1536]> linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_mlp_up_proj_weight_to_fp16, x = x_157_cast_fp16)[name = tensor<string, []>("linear_26_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_159_cast_fp16 = mul(x = var_815_cast_fp16, y = linear_26_cast_fp16)[name = tensor<string, []>("x_159_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_3_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_3_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256550144)))];
            tensor<fp16, [1, 1, 512]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_3_mlp_down_proj_weight_to_fp16, x = x_159_cast_fp16)[name = tensor<string, []>("linear_27_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_161_cast_fp16 = add(x = x_149_cast_fp16, y = linear_27_cast_fp16)[name = tensor<string, []>("x_161_cast_fp16")];
            tensor<string, []> x_161_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_161_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_9_begin_0 = const()[name = tensor<string, []>("k_cache_9_begin_0"), val = tensor<int32, [5]>([4, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_9_end_0 = const()[name = tensor<string, []>("k_cache_9_end_0"), val = tensor<int32, [5]>([5, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_9_end_mask_0 = const()[name = tensor<string, []>("k_cache_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_9_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_9_cast_fp16 = slice_by_index(begin = k_cache_9_begin_0, end = k_cache_9_end_0, end_mask = k_cache_9_end_mask_0, squeeze_mask = k_cache_9_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_9_cast_fp16")];
            tensor<int32, [5]> v_cache_9_begin_0 = const()[name = tensor<string, []>("v_cache_9_begin_0"), val = tensor<int32, [5]>([4, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_9_end_0 = const()[name = tensor<string, []>("v_cache_9_end_0"), val = tensor<int32, [5]>([5, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_9_end_mask_0 = const()[name = tensor<string, []>("v_cache_9_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_9_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_9_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_9_cast_fp16 = slice_by_index(begin = v_cache_9_begin_0, end = v_cache_9_end_0, end_mask = v_cache_9_end_mask_0, squeeze_mask = v_cache_9_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_9_cast_fp16")];
            tensor<int32, []> var_847 = const()[name = tensor<string, []>("op_847"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_846_promoted = const()[name = tensor<string, []>("op_846_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_161_cast_fp16_to_fp32 = cast(dtype = x_161_cast_fp16_to_fp32_dtype_0, x = x_161_cast_fp16)[name = tensor<string, []>("cast_614")];
            tensor<fp32, [1, 1, 512]> var_856 = pow(x = x_161_cast_fp16_to_fp32, y = var_846_promoted)[name = tensor<string, []>("op_856")];
            tensor<int32, [1]> var_33_axes_0 = const()[name = tensor<string, []>("var_33_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_33_keep_dims_0 = const()[name = tensor<string, []>("var_33_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_33 = reduce_mean(axes = var_33_axes_0, keep_dims = var_33_keep_dims_0, x = var_856)[name = tensor<string, []>("var_33")];
            tensor<string, []> var_33_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_33_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_860_to_fp16 = const()[name = tensor<string, []>("op_860_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_33_to_fp16 = cast(dtype = var_33_to_fp16_dtype_0, x = var_33)[name = tensor<string, []>("cast_613")];
            tensor<fp16, [1, 1, 1]> var_861_cast_fp16 = add(x = var_33_to_fp16, y = var_860_to_fp16)[name = tensor<string, []>("op_861_cast_fp16")];
            tensor<string, []> var_861_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_861_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_862_epsilon_0 = const()[name = tensor<string, []>("op_862_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_861_cast_fp16_to_fp32 = cast(dtype = var_861_cast_fp16_to_fp32_dtype_0, x = var_861_cast_fp16)[name = tensor<string, []>("cast_612")];
            tensor<fp32, [1, 1, 1]> var_862 = rsqrt(epsilon = var_862_epsilon_0, x = var_861_cast_fp16_to_fp32)[name = tensor<string, []>("op_862")];
            tensor<string, []> var_862_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_862_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_862_to_fp16 = cast(dtype = var_862_to_fp16_dtype_0, x = var_862)[name = tensor<string, []>("cast_611")];
            tensor<fp16, [1, 1, 512]> x_167_cast_fp16 = mul(x = x_161_cast_fp16, y = var_862_to_fp16)[name = tensor<string, []>("x_167_cast_fp16")];
            tensor<fp16, [512]> layers_4_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258123072)))];
            tensor<fp16, [1, 1, 512]> x_169_cast_fp16 = mul(x = layers_4_input_layernorm_weight_to_fp16, y = x_167_cast_fp16)[name = tensor<string, []>("x_169_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(258124160)))];
            tensor<fp16, [1, 1, 1536]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = x_169_cast_fp16)[name = tensor<string, []>("linear_28_cast_fp16")];
            tensor<int32, [4]> var_876 = const()[name = tensor<string, []>("op_876"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_171_cast_fp16 = reshape(shape = var_876, x = linear_28_cast_fp16)[name = tensor<string, []>("x_171_cast_fp16")];
            tensor<string, []> x_171_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_171_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_846_promoted_1 = const()[name = tensor<string, []>("op_846_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_171_cast_fp16_to_fp32 = cast(dtype = x_171_cast_fp16_to_fp32_dtype_0, x = x_171_cast_fp16)[name = tensor<string, []>("cast_610")];
            tensor<fp32, [1, 1, 12, 128]> var_880 = pow(x = x_171_cast_fp16_to_fp32, y = var_846_promoted_1)[name = tensor<string, []>("op_880")];
            tensor<int32, [1]> var_35_axes_0 = const()[name = tensor<string, []>("var_35_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_35_keep_dims_0 = const()[name = tensor<string, []>("var_35_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_35 = reduce_mean(axes = var_35_axes_0, keep_dims = var_35_keep_dims_0, x = var_880)[name = tensor<string, []>("var_35")];
            tensor<string, []> var_35_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_35_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_884_to_fp16 = const()[name = tensor<string, []>("op_884_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_35_to_fp16 = cast(dtype = var_35_to_fp16_dtype_0, x = var_35)[name = tensor<string, []>("cast_609")];
            tensor<fp16, [1, 1, 12, 1]> var_885_cast_fp16 = add(x = var_35_to_fp16, y = var_884_to_fp16)[name = tensor<string, []>("op_885_cast_fp16")];
            tensor<string, []> var_885_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_885_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_886_epsilon_0 = const()[name = tensor<string, []>("op_886_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_885_cast_fp16_to_fp32 = cast(dtype = var_885_cast_fp16_to_fp32_dtype_0, x = var_885_cast_fp16)[name = tensor<string, []>("cast_608")];
            tensor<fp32, [1, 1, 12, 1]> var_886 = rsqrt(epsilon = var_886_epsilon_0, x = var_885_cast_fp16_to_fp32)[name = tensor<string, []>("op_886")];
            tensor<string, []> var_886_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_886_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_886_to_fp16 = cast(dtype = var_886_to_fp16_dtype_0, x = var_886)[name = tensor<string, []>("cast_607")];
            tensor<fp16, [1, 1, 12, 128]> x_177_cast_fp16 = mul(x = x_171_cast_fp16, y = var_886_to_fp16)[name = tensor<string, []>("x_177_cast_fp16")];
            tensor<fp16, [128]> layers_4_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259697088)))];
            tensor<fp16, [1, 1, 12, 128]> var_888_cast_fp16 = mul(x = layers_4_self_attn_q_norm_weight_to_fp16, y = x_177_cast_fp16)[name = tensor<string, []>("op_888_cast_fp16")];
            tensor<int32, [4]> q_17_perm_0 = const()[name = tensor<string, []>("q_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259697408)))];
            tensor<fp16, [1, 1, 512]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = x_169_cast_fp16)[name = tensor<string, []>("linear_29_cast_fp16")];
            tensor<int32, [4]> var_892 = const()[name = tensor<string, []>("op_892"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_179_cast_fp16 = reshape(shape = var_892, x = linear_29_cast_fp16)[name = tensor<string, []>("x_179_cast_fp16")];
            tensor<string, []> x_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_179_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_846_promoted_2 = const()[name = tensor<string, []>("op_846_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_179_cast_fp16_to_fp32 = cast(dtype = x_179_cast_fp16_to_fp32_dtype_0, x = x_179_cast_fp16)[name = tensor<string, []>("cast_606")];
            tensor<fp32, [1, 1, 4, 128]> var_896 = pow(x = x_179_cast_fp16_to_fp32, y = var_846_promoted_2)[name = tensor<string, []>("op_896")];
            tensor<int32, [1]> var_37_axes_0 = const()[name = tensor<string, []>("var_37_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_37_keep_dims_0 = const()[name = tensor<string, []>("var_37_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_37 = reduce_mean(axes = var_37_axes_0, keep_dims = var_37_keep_dims_0, x = var_896)[name = tensor<string, []>("var_37")];
            tensor<string, []> var_37_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_37_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_900_to_fp16 = const()[name = tensor<string, []>("op_900_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_37_to_fp16 = cast(dtype = var_37_to_fp16_dtype_0, x = var_37)[name = tensor<string, []>("cast_605")];
            tensor<fp16, [1, 1, 4, 1]> var_901_cast_fp16 = add(x = var_37_to_fp16, y = var_900_to_fp16)[name = tensor<string, []>("op_901_cast_fp16")];
            tensor<string, []> var_901_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_901_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_902_epsilon_0 = const()[name = tensor<string, []>("op_902_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_901_cast_fp16_to_fp32 = cast(dtype = var_901_cast_fp16_to_fp32_dtype_0, x = var_901_cast_fp16)[name = tensor<string, []>("cast_604")];
            tensor<fp32, [1, 1, 4, 1]> var_902 = rsqrt(epsilon = var_902_epsilon_0, x = var_901_cast_fp16_to_fp32)[name = tensor<string, []>("op_902")];
            tensor<string, []> var_902_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_902_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_902_to_fp16 = cast(dtype = var_902_to_fp16_dtype_0, x = var_902)[name = tensor<string, []>("cast_603")];
            tensor<fp16, [1, 1, 4, 128]> x_185_cast_fp16 = mul(x = x_179_cast_fp16, y = var_902_to_fp16)[name = tensor<string, []>("x_185_cast_fp16")];
            tensor<fp16, [128]> layers_4_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260221760)))];
            tensor<fp16, [1, 1, 4, 128]> var_904_cast_fp16 = mul(x = layers_4_self_attn_k_norm_weight_to_fp16, y = x_185_cast_fp16)[name = tensor<string, []>("op_904_cast_fp16")];
            tensor<int32, [4]> k_17_perm_0 = const()[name = tensor<string, []>("k_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260222080)))];
            tensor<fp16, [1, 1, 512]> linear_30_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = x_169_cast_fp16)[name = tensor<string, []>("linear_30_cast_fp16")];
            tensor<int32, [4]> var_908 = const()[name = tensor<string, []>("op_908"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_909_cast_fp16 = reshape(shape = var_908, x = linear_30_cast_fp16)[name = tensor<string, []>("op_909_cast_fp16")];
            tensor<int32, [4]> v_9_perm_0 = const()[name = tensor<string, []>("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_17_cast_fp16 = transpose(perm = q_17_perm_0, x = var_888_cast_fp16)[name = tensor<string, []>("transpose_95")];
            tensor<fp16, [1, 12, 1, 128]> var_913_cast_fp16 = mul(x = q_17_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_913_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = tensor<string, []>("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = tensor<string, []>("x1_17_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = tensor<string, []>("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_17_cast_fp16)[name = tensor<string, []>("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = tensor<string, []>("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = tensor<string, []>("x2_17_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = tensor<string, []>("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_17_cast_fp16)[name = tensor<string, []>("x2_17_cast_fp16")];
            tensor<fp16, []> const_9_promoted_to_fp16 = const()[name = tensor<string, []>("const_9_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_916_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_9_promoted_to_fp16)[name = tensor<string, []>("op_916_cast_fp16")];
            tensor<bool, []> var_918_interleave_0 = const()[name = tensor<string, []>("op_918_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_918_cast_fp16 = concat(axis = var_847, interleave = var_918_interleave_0, values = (var_916_cast_fp16, x1_17_cast_fp16))[name = tensor<string, []>("op_918_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_919_cast_fp16 = mul(x = var_918_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_919_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_19_cast_fp16 = add(x = var_913_cast_fp16, y = var_919_cast_fp16)[name = tensor<string, []>("q_19_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_17_cast_fp16 = transpose(perm = k_17_perm_0, x = var_904_cast_fp16)[name = tensor<string, []>("transpose_94")];
            tensor<fp16, [1, 4, 1, 128]> var_921_cast_fp16 = mul(x = k_17_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_921_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = tensor<string, []>("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = tensor<string, []>("x1_19_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = tensor<string, []>("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_17_cast_fp16)[name = tensor<string, []>("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = tensor<string, []>("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = tensor<string, []>("x2_19_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = tensor<string, []>("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_17_cast_fp16)[name = tensor<string, []>("x2_19_cast_fp16")];
            tensor<fp16, []> const_10_promoted_to_fp16 = const()[name = tensor<string, []>("const_10_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_924_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_10_promoted_to_fp16)[name = tensor<string, []>("op_924_cast_fp16")];
            tensor<bool, []> var_926_interleave_0 = const()[name = tensor<string, []>("op_926_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_926_cast_fp16 = concat(axis = var_847, interleave = var_926_interleave_0, values = (var_924_cast_fp16, x1_19_cast_fp16))[name = tensor<string, []>("op_926_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_927_cast_fp16 = mul(x = var_926_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_927_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_19_cast_fp16 = add(x = var_921_cast_fp16, y = var_927_cast_fp16)[name = tensor<string, []>("k_19_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_930_cast_fp16 = mul(x = k_cache_9_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_930_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_931_cast_fp16 = mul(x = k_19_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_931_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_9_cast_fp16 = add(x = var_930_cast_fp16, y = var_931_cast_fp16)[name = tensor<string, []>("k_full_9_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_934_cast_fp16 = mul(x = v_cache_9_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_934_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_909_cast_fp16)[name = tensor<string, []>("transpose_93")];
            tensor<fp16, [1, 4, 2048, 128]> var_935_cast_fp16 = mul(x = v_9_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_935_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_9_cast_fp16 = add(x = var_934_cast_fp16, y = var_935_cast_fp16)[name = tensor<string, []>("v_full_9_cast_fp16")];
            tensor<int32, [1]> var_937_axes_0 = const()[name = tensor<string, []>("op_937_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_937_cast_fp16 = expand_dims(axes = var_937_axes_0, x = k_full_9_cast_fp16)[name = tensor<string, []>("op_937_cast_fp16")];
            tensor<int32, [5]> var_939_reps_0 = const()[name = tensor<string, []>("op_939_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_939_cast_fp16 = tile(reps = var_939_reps_0, x = var_937_cast_fp16)[name = tensor<string, []>("op_939_cast_fp16")];
            tensor<int32, [4]> var_940 = const()[name = tensor<string, []>("op_940"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_9_cast_fp16 = reshape(shape = var_940, x = var_939_cast_fp16)[name = tensor<string, []>("k_rep_9_cast_fp16")];
            tensor<int32, [1]> var_942_axes_0 = const()[name = tensor<string, []>("op_942_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_942_cast_fp16 = expand_dims(axes = var_942_axes_0, x = v_full_9_cast_fp16)[name = tensor<string, []>("op_942_cast_fp16")];
            tensor<int32, [5]> var_944_reps_0 = const()[name = tensor<string, []>("op_944_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_944_cast_fp16 = tile(reps = var_944_reps_0, x = var_942_cast_fp16)[name = tensor<string, []>("op_944_cast_fp16")];
            tensor<int32, [4]> var_945 = const()[name = tensor<string, []>("op_945"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_9_cast_fp16 = reshape(shape = var_945, x = var_944_cast_fp16)[name = tensor<string, []>("v_rep_9_cast_fp16")];
            tensor<bool, []> var_948_transpose_x_1 = const()[name = tensor<string, []>("op_948_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_948_transpose_y_1 = const()[name = tensor<string, []>("op_948_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_948_cast_fp16 = matmul(transpose_x = var_948_transpose_x_1, transpose_y = var_948_transpose_y_1, x = q_19_cast_fp16, y = k_rep_9_cast_fp16)[name = tensor<string, []>("op_948_cast_fp16")];
            tensor<fp16, []> var_949_to_fp16 = const()[name = tensor<string, []>("op_949_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_17_cast_fp16 = mul(x = var_948_cast_fp16, y = var_949_to_fp16)[name = tensor<string, []>("attn_17_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_19_cast_fp16 = add(x = attn_17_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
            tensor<string, []> input_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_19_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_19_cast_fp16_to_fp32 = cast(dtype = input_19_cast_fp16_to_fp32_dtype_0, x = input_19_cast_fp16)[name = tensor<string, []>("cast_602")];
            tensor<fp32, [1, 12, 1, 2048]> attn_19 = softmax(axis = var_847, x = input_19_cast_fp16_to_fp32)[name = tensor<string, []>("attn_19")];
            tensor<bool, []> out_9_transpose_x_0 = const()[name = tensor<string, []>("out_9_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_9_transpose_y_0 = const()[name = tensor<string, []>("out_9_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_19_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_19_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_19_to_fp16 = cast(dtype = attn_19_to_fp16_dtype_0, x = attn_19)[name = tensor<string, []>("cast_601")];
            tensor<fp16, [1, 12, 1, 128]> out_9_cast_fp16 = matmul(transpose_x = out_9_transpose_x_0, transpose_y = out_9_transpose_y_0, x = attn_19_to_fp16, y = v_rep_9_cast_fp16)[name = tensor<string, []>("out_9_cast_fp16")];
            tensor<int32, [4]> var_954_perm_0 = const()[name = tensor<string, []>("op_954_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_956 = const()[name = tensor<string, []>("op_956"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_954_cast_fp16 = transpose(perm = var_954_perm_0, x = out_9_cast_fp16)[name = tensor<string, []>("transpose_92")];
            tensor<fp16, [1, 1, 1536]> x_187_cast_fp16 = reshape(shape = var_956, x = var_954_cast_fp16)[name = tensor<string, []>("x_187_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260746432)))];
            tensor<fp16, [1, 1, 512]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = x_187_cast_fp16)[name = tensor<string, []>("linear_31_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_189_cast_fp16 = add(x = x_161_cast_fp16, y = linear_31_cast_fp16)[name = tensor<string, []>("x_189_cast_fp16")];
            tensor<string, []> x_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_189_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_846_promoted_3 = const()[name = tensor<string, []>("op_846_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_189_cast_fp16_to_fp32 = cast(dtype = x_189_cast_fp16_to_fp32_dtype_0, x = x_189_cast_fp16)[name = tensor<string, []>("cast_600")];
            tensor<fp32, [1, 1, 512]> var_967 = pow(x = x_189_cast_fp16_to_fp32, y = var_846_promoted_3)[name = tensor<string, []>("op_967")];
            tensor<int32, [1]> var_39_axes_0 = const()[name = tensor<string, []>("var_39_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_39_keep_dims_0 = const()[name = tensor<string, []>("var_39_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_39 = reduce_mean(axes = var_39_axes_0, keep_dims = var_39_keep_dims_0, x = var_967)[name = tensor<string, []>("var_39")];
            tensor<string, []> var_39_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_39_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_971_to_fp16 = const()[name = tensor<string, []>("op_971_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_39_to_fp16 = cast(dtype = var_39_to_fp16_dtype_0, x = var_39)[name = tensor<string, []>("cast_599")];
            tensor<fp16, [1, 1, 1]> var_972_cast_fp16 = add(x = var_39_to_fp16, y = var_971_to_fp16)[name = tensor<string, []>("op_972_cast_fp16")];
            tensor<string, []> var_972_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_972_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_973_epsilon_0 = const()[name = tensor<string, []>("op_973_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_972_cast_fp16_to_fp32 = cast(dtype = var_972_cast_fp16_to_fp32_dtype_0, x = var_972_cast_fp16)[name = tensor<string, []>("cast_598")];
            tensor<fp32, [1, 1, 1]> var_973 = rsqrt(epsilon = var_973_epsilon_0, x = var_972_cast_fp16_to_fp32)[name = tensor<string, []>("op_973")];
            tensor<string, []> var_973_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_973_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_973_to_fp16 = cast(dtype = var_973_to_fp16_dtype_0, x = var_973)[name = tensor<string, []>("cast_597")];
            tensor<fp16, [1, 1, 512]> x_195_cast_fp16 = mul(x = x_189_cast_fp16, y = var_973_to_fp16)[name = tensor<string, []>("x_195_cast_fp16")];
            tensor<fp16, [512]> layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262319360)))];
            tensor<fp16, [1, 1, 512]> x_197_cast_fp16 = mul(x = layers_4_post_attention_layernorm_weight_to_fp16, y = x_195_cast_fp16)[name = tensor<string, []>("x_197_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_4_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262320448)))];
            tensor<fp16, [1, 1, 1536]> linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_mlp_gate_proj_weight_to_fp16, x = x_197_cast_fp16)[name = tensor<string, []>("linear_32_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_984_cast_fp16 = silu(x = linear_32_cast_fp16)[name = tensor<string, []>("op_984_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_4_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263893376)))];
            tensor<fp16, [1, 1, 1536]> linear_33_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_mlp_up_proj_weight_to_fp16, x = x_197_cast_fp16)[name = tensor<string, []>("linear_33_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_199_cast_fp16 = mul(x = var_984_cast_fp16, y = linear_33_cast_fp16)[name = tensor<string, []>("x_199_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_4_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_4_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(265466304)))];
            tensor<fp16, [1, 1, 512]> linear_34_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_4_mlp_down_proj_weight_to_fp16, x = x_199_cast_fp16)[name = tensor<string, []>("linear_34_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_201_cast_fp16 = add(x = x_189_cast_fp16, y = linear_34_cast_fp16)[name = tensor<string, []>("x_201_cast_fp16")];
            tensor<string, []> x_201_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_201_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_11_begin_0 = const()[name = tensor<string, []>("k_cache_11_begin_0"), val = tensor<int32, [5]>([5, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_11_end_0 = const()[name = tensor<string, []>("k_cache_11_end_0"), val = tensor<int32, [5]>([6, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_11_end_mask_0 = const()[name = tensor<string, []>("k_cache_11_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_11_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_11_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_11_cast_fp16 = slice_by_index(begin = k_cache_11_begin_0, end = k_cache_11_end_0, end_mask = k_cache_11_end_mask_0, squeeze_mask = k_cache_11_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_11_cast_fp16")];
            tensor<int32, [5]> v_cache_11_begin_0 = const()[name = tensor<string, []>("v_cache_11_begin_0"), val = tensor<int32, [5]>([5, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_11_end_0 = const()[name = tensor<string, []>("v_cache_11_end_0"), val = tensor<int32, [5]>([6, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_11_end_mask_0 = const()[name = tensor<string, []>("v_cache_11_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_11_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_11_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_11_cast_fp16 = slice_by_index(begin = v_cache_11_begin_0, end = v_cache_11_end_0, end_mask = v_cache_11_end_mask_0, squeeze_mask = v_cache_11_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_11_cast_fp16")];
            tensor<int32, []> var_1016 = const()[name = tensor<string, []>("op_1016"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_1015_promoted = const()[name = tensor<string, []>("op_1015_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_201_cast_fp16_to_fp32 = cast(dtype = x_201_cast_fp16_to_fp32_dtype_0, x = x_201_cast_fp16)[name = tensor<string, []>("cast_596")];
            tensor<fp32, [1, 1, 512]> var_1025 = pow(x = x_201_cast_fp16_to_fp32, y = var_1015_promoted)[name = tensor<string, []>("op_1025")];
            tensor<int32, [1]> var_41_axes_0 = const()[name = tensor<string, []>("var_41_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_41_keep_dims_0 = const()[name = tensor<string, []>("var_41_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_41 = reduce_mean(axes = var_41_axes_0, keep_dims = var_41_keep_dims_0, x = var_1025)[name = tensor<string, []>("var_41")];
            tensor<string, []> var_41_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_41_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1029_to_fp16 = const()[name = tensor<string, []>("op_1029_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_41_to_fp16 = cast(dtype = var_41_to_fp16_dtype_0, x = var_41)[name = tensor<string, []>("cast_595")];
            tensor<fp16, [1, 1, 1]> var_1030_cast_fp16 = add(x = var_41_to_fp16, y = var_1029_to_fp16)[name = tensor<string, []>("op_1030_cast_fp16")];
            tensor<string, []> var_1030_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1030_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1031_epsilon_0 = const()[name = tensor<string, []>("op_1031_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1030_cast_fp16_to_fp32 = cast(dtype = var_1030_cast_fp16_to_fp32_dtype_0, x = var_1030_cast_fp16)[name = tensor<string, []>("cast_594")];
            tensor<fp32, [1, 1, 1]> var_1031 = rsqrt(epsilon = var_1031_epsilon_0, x = var_1030_cast_fp16_to_fp32)[name = tensor<string, []>("op_1031")];
            tensor<string, []> var_1031_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1031_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1031_to_fp16 = cast(dtype = var_1031_to_fp16_dtype_0, x = var_1031)[name = tensor<string, []>("cast_593")];
            tensor<fp16, [1, 1, 512]> x_207_cast_fp16 = mul(x = x_201_cast_fp16, y = var_1031_to_fp16)[name = tensor<string, []>("x_207_cast_fp16")];
            tensor<fp16, [512]> layers_5_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267039232)))];
            tensor<fp16, [1, 1, 512]> x_209_cast_fp16 = mul(x = layers_5_input_layernorm_weight_to_fp16, y = x_207_cast_fp16)[name = tensor<string, []>("x_209_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267040320)))];
            tensor<fp16, [1, 1, 1536]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = x_209_cast_fp16)[name = tensor<string, []>("linear_35_cast_fp16")];
            tensor<int32, [4]> var_1045 = const()[name = tensor<string, []>("op_1045"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_211_cast_fp16 = reshape(shape = var_1045, x = linear_35_cast_fp16)[name = tensor<string, []>("x_211_cast_fp16")];
            tensor<string, []> x_211_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_211_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1015_promoted_1 = const()[name = tensor<string, []>("op_1015_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_211_cast_fp16_to_fp32 = cast(dtype = x_211_cast_fp16_to_fp32_dtype_0, x = x_211_cast_fp16)[name = tensor<string, []>("cast_592")];
            tensor<fp32, [1, 1, 12, 128]> var_1049 = pow(x = x_211_cast_fp16_to_fp32, y = var_1015_promoted_1)[name = tensor<string, []>("op_1049")];
            tensor<int32, [1]> var_43_axes_0 = const()[name = tensor<string, []>("var_43_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_43_keep_dims_0 = const()[name = tensor<string, []>("var_43_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_43 = reduce_mean(axes = var_43_axes_0, keep_dims = var_43_keep_dims_0, x = var_1049)[name = tensor<string, []>("var_43")];
            tensor<string, []> var_43_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_43_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1053_to_fp16 = const()[name = tensor<string, []>("op_1053_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_43_to_fp16 = cast(dtype = var_43_to_fp16_dtype_0, x = var_43)[name = tensor<string, []>("cast_591")];
            tensor<fp16, [1, 1, 12, 1]> var_1054_cast_fp16 = add(x = var_43_to_fp16, y = var_1053_to_fp16)[name = tensor<string, []>("op_1054_cast_fp16")];
            tensor<string, []> var_1054_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1054_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1055_epsilon_0 = const()[name = tensor<string, []>("op_1055_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_1054_cast_fp16_to_fp32 = cast(dtype = var_1054_cast_fp16_to_fp32_dtype_0, x = var_1054_cast_fp16)[name = tensor<string, []>("cast_590")];
            tensor<fp32, [1, 1, 12, 1]> var_1055 = rsqrt(epsilon = var_1055_epsilon_0, x = var_1054_cast_fp16_to_fp32)[name = tensor<string, []>("op_1055")];
            tensor<string, []> var_1055_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1055_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_1055_to_fp16 = cast(dtype = var_1055_to_fp16_dtype_0, x = var_1055)[name = tensor<string, []>("cast_589")];
            tensor<fp16, [1, 1, 12, 128]> x_217_cast_fp16 = mul(x = x_211_cast_fp16, y = var_1055_to_fp16)[name = tensor<string, []>("x_217_cast_fp16")];
            tensor<fp16, [128]> layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268613248)))];
            tensor<fp16, [1, 1, 12, 128]> var_1057_cast_fp16 = mul(x = layers_5_self_attn_q_norm_weight_to_fp16, y = x_217_cast_fp16)[name = tensor<string, []>("op_1057_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = tensor<string, []>("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(268613568)))];
            tensor<fp16, [1, 1, 512]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = x_209_cast_fp16)[name = tensor<string, []>("linear_36_cast_fp16")];
            tensor<int32, [4]> var_1061 = const()[name = tensor<string, []>("op_1061"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_219_cast_fp16 = reshape(shape = var_1061, x = linear_36_cast_fp16)[name = tensor<string, []>("x_219_cast_fp16")];
            tensor<string, []> x_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_219_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1015_promoted_2 = const()[name = tensor<string, []>("op_1015_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_219_cast_fp16_to_fp32 = cast(dtype = x_219_cast_fp16_to_fp32_dtype_0, x = x_219_cast_fp16)[name = tensor<string, []>("cast_588")];
            tensor<fp32, [1, 1, 4, 128]> var_1065 = pow(x = x_219_cast_fp16_to_fp32, y = var_1015_promoted_2)[name = tensor<string, []>("op_1065")];
            tensor<int32, [1]> var_45_axes_0 = const()[name = tensor<string, []>("var_45_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_45_keep_dims_0 = const()[name = tensor<string, []>("var_45_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_45 = reduce_mean(axes = var_45_axes_0, keep_dims = var_45_keep_dims_0, x = var_1065)[name = tensor<string, []>("var_45")];
            tensor<string, []> var_45_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_45_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1069_to_fp16 = const()[name = tensor<string, []>("op_1069_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_45_to_fp16 = cast(dtype = var_45_to_fp16_dtype_0, x = var_45)[name = tensor<string, []>("cast_587")];
            tensor<fp16, [1, 1, 4, 1]> var_1070_cast_fp16 = add(x = var_45_to_fp16, y = var_1069_to_fp16)[name = tensor<string, []>("op_1070_cast_fp16")];
            tensor<string, []> var_1070_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1070_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1071_epsilon_0 = const()[name = tensor<string, []>("op_1071_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_1070_cast_fp16_to_fp32 = cast(dtype = var_1070_cast_fp16_to_fp32_dtype_0, x = var_1070_cast_fp16)[name = tensor<string, []>("cast_586")];
            tensor<fp32, [1, 1, 4, 1]> var_1071 = rsqrt(epsilon = var_1071_epsilon_0, x = var_1070_cast_fp16_to_fp32)[name = tensor<string, []>("op_1071")];
            tensor<string, []> var_1071_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1071_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_1071_to_fp16 = cast(dtype = var_1071_to_fp16_dtype_0, x = var_1071)[name = tensor<string, []>("cast_585")];
            tensor<fp16, [1, 1, 4, 128]> x_225_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1071_to_fp16)[name = tensor<string, []>("x_225_cast_fp16")];
            tensor<fp16, [128]> layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269137920)))];
            tensor<fp16, [1, 1, 4, 128]> var_1073_cast_fp16 = mul(x = layers_5_self_attn_k_norm_weight_to_fp16, y = x_225_cast_fp16)[name = tensor<string, []>("op_1073_cast_fp16")];
            tensor<int32, [4]> k_21_perm_0 = const()[name = tensor<string, []>("k_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269138240)))];
            tensor<fp16, [1, 1, 512]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = x_209_cast_fp16)[name = tensor<string, []>("linear_37_cast_fp16")];
            tensor<int32, [4]> var_1077 = const()[name = tensor<string, []>("op_1077"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_1078_cast_fp16 = reshape(shape = var_1077, x = linear_37_cast_fp16)[name = tensor<string, []>("op_1078_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = tensor<string, []>("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = var_1057_cast_fp16)[name = tensor<string, []>("transpose_91")];
            tensor<fp16, [1, 12, 1, 128]> var_1082_cast_fp16 = mul(x = q_21_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1082_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = tensor<string, []>("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = tensor<string, []>("x1_21_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = tensor<string, []>("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_21_cast_fp16)[name = tensor<string, []>("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = tensor<string, []>("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = tensor<string, []>("x2_21_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = tensor<string, []>("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_21_cast_fp16)[name = tensor<string, []>("x2_21_cast_fp16")];
            tensor<fp16, []> const_11_promoted_to_fp16 = const()[name = tensor<string, []>("const_11_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_1085_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_11_promoted_to_fp16)[name = tensor<string, []>("op_1085_cast_fp16")];
            tensor<bool, []> var_1087_interleave_0 = const()[name = tensor<string, []>("op_1087_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_1087_cast_fp16 = concat(axis = var_1016, interleave = var_1087_interleave_0, values = (var_1085_cast_fp16, x1_21_cast_fp16))[name = tensor<string, []>("op_1087_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_1088_cast_fp16 = mul(x = var_1087_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1088_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_23_cast_fp16 = add(x = var_1082_cast_fp16, y = var_1088_cast_fp16)[name = tensor<string, []>("q_23_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_21_cast_fp16 = transpose(perm = k_21_perm_0, x = var_1073_cast_fp16)[name = tensor<string, []>("transpose_90")];
            tensor<fp16, [1, 4, 1, 128]> var_1090_cast_fp16 = mul(x = k_21_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1090_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = tensor<string, []>("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = tensor<string, []>("x1_23_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = tensor<string, []>("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_21_cast_fp16)[name = tensor<string, []>("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = tensor<string, []>("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = tensor<string, []>("x2_23_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = tensor<string, []>("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_21_cast_fp16)[name = tensor<string, []>("x2_23_cast_fp16")];
            tensor<fp16, []> const_12_promoted_to_fp16 = const()[name = tensor<string, []>("const_12_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_1093_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_12_promoted_to_fp16)[name = tensor<string, []>("op_1093_cast_fp16")];
            tensor<bool, []> var_1095_interleave_0 = const()[name = tensor<string, []>("op_1095_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_1095_cast_fp16 = concat(axis = var_1016, interleave = var_1095_interleave_0, values = (var_1093_cast_fp16, x1_23_cast_fp16))[name = tensor<string, []>("op_1095_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_1096_cast_fp16 = mul(x = var_1095_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1096_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_23_cast_fp16 = add(x = var_1090_cast_fp16, y = var_1096_cast_fp16)[name = tensor<string, []>("k_23_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1099_cast_fp16 = mul(x = k_cache_11_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1099_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1100_cast_fp16 = mul(x = k_23_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1100_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_11_cast_fp16 = add(x = var_1099_cast_fp16, y = var_1100_cast_fp16)[name = tensor<string, []>("k_full_11_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1103_cast_fp16 = mul(x = v_cache_11_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1103_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_1078_cast_fp16)[name = tensor<string, []>("transpose_89")];
            tensor<fp16, [1, 4, 2048, 128]> var_1104_cast_fp16 = mul(x = v_11_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1104_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_11_cast_fp16 = add(x = var_1103_cast_fp16, y = var_1104_cast_fp16)[name = tensor<string, []>("v_full_11_cast_fp16")];
            tensor<int32, [1]> var_1106_axes_0 = const()[name = tensor<string, []>("op_1106_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1106_cast_fp16 = expand_dims(axes = var_1106_axes_0, x = k_full_11_cast_fp16)[name = tensor<string, []>("op_1106_cast_fp16")];
            tensor<int32, [5]> var_1108_reps_0 = const()[name = tensor<string, []>("op_1108_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1108_cast_fp16 = tile(reps = var_1108_reps_0, x = var_1106_cast_fp16)[name = tensor<string, []>("op_1108_cast_fp16")];
            tensor<int32, [4]> var_1109 = const()[name = tensor<string, []>("op_1109"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_11_cast_fp16 = reshape(shape = var_1109, x = var_1108_cast_fp16)[name = tensor<string, []>("k_rep_11_cast_fp16")];
            tensor<int32, [1]> var_1111_axes_0 = const()[name = tensor<string, []>("op_1111_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1111_cast_fp16 = expand_dims(axes = var_1111_axes_0, x = v_full_11_cast_fp16)[name = tensor<string, []>("op_1111_cast_fp16")];
            tensor<int32, [5]> var_1113_reps_0 = const()[name = tensor<string, []>("op_1113_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1113_cast_fp16 = tile(reps = var_1113_reps_0, x = var_1111_cast_fp16)[name = tensor<string, []>("op_1113_cast_fp16")];
            tensor<int32, [4]> var_1114 = const()[name = tensor<string, []>("op_1114"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_11_cast_fp16 = reshape(shape = var_1114, x = var_1113_cast_fp16)[name = tensor<string, []>("v_rep_11_cast_fp16")];
            tensor<bool, []> var_1117_transpose_x_1 = const()[name = tensor<string, []>("op_1117_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1117_transpose_y_1 = const()[name = tensor<string, []>("op_1117_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_1117_cast_fp16 = matmul(transpose_x = var_1117_transpose_x_1, transpose_y = var_1117_transpose_y_1, x = q_23_cast_fp16, y = k_rep_11_cast_fp16)[name = tensor<string, []>("op_1117_cast_fp16")];
            tensor<fp16, []> var_1118_to_fp16 = const()[name = tensor<string, []>("op_1118_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_21_cast_fp16 = mul(x = var_1117_cast_fp16, y = var_1118_to_fp16)[name = tensor<string, []>("attn_21_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_23_cast_fp16 = add(x = attn_21_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_23_cast_fp16")];
            tensor<string, []> input_23_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_23_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_23_cast_fp16_to_fp32 = cast(dtype = input_23_cast_fp16_to_fp32_dtype_0, x = input_23_cast_fp16)[name = tensor<string, []>("cast_584")];
            tensor<fp32, [1, 12, 1, 2048]> attn_23 = softmax(axis = var_1016, x = input_23_cast_fp16_to_fp32)[name = tensor<string, []>("attn_23")];
            tensor<bool, []> out_11_transpose_x_0 = const()[name = tensor<string, []>("out_11_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_11_transpose_y_0 = const()[name = tensor<string, []>("out_11_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_23_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_23_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_23_to_fp16 = cast(dtype = attn_23_to_fp16_dtype_0, x = attn_23)[name = tensor<string, []>("cast_583")];
            tensor<fp16, [1, 12, 1, 128]> out_11_cast_fp16 = matmul(transpose_x = out_11_transpose_x_0, transpose_y = out_11_transpose_y_0, x = attn_23_to_fp16, y = v_rep_11_cast_fp16)[name = tensor<string, []>("out_11_cast_fp16")];
            tensor<int32, [4]> var_1123_perm_0 = const()[name = tensor<string, []>("op_1123_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1125 = const()[name = tensor<string, []>("op_1125"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_1123_cast_fp16 = transpose(perm = var_1123_perm_0, x = out_11_cast_fp16)[name = tensor<string, []>("transpose_88")];
            tensor<fp16, [1, 1, 1536]> x_227_cast_fp16 = reshape(shape = var_1125, x = var_1123_cast_fp16)[name = tensor<string, []>("x_227_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(269662592)))];
            tensor<fp16, [1, 1, 512]> linear_38_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = x_227_cast_fp16)[name = tensor<string, []>("linear_38_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_229_cast_fp16 = add(x = x_201_cast_fp16, y = linear_38_cast_fp16)[name = tensor<string, []>("x_229_cast_fp16")];
            tensor<string, []> x_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_229_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1015_promoted_3 = const()[name = tensor<string, []>("op_1015_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_229_cast_fp16_to_fp32 = cast(dtype = x_229_cast_fp16_to_fp32_dtype_0, x = x_229_cast_fp16)[name = tensor<string, []>("cast_582")];
            tensor<fp32, [1, 1, 512]> var_1136 = pow(x = x_229_cast_fp16_to_fp32, y = var_1015_promoted_3)[name = tensor<string, []>("op_1136")];
            tensor<int32, [1]> var_47_axes_0 = const()[name = tensor<string, []>("var_47_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_47_keep_dims_0 = const()[name = tensor<string, []>("var_47_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_47 = reduce_mean(axes = var_47_axes_0, keep_dims = var_47_keep_dims_0, x = var_1136)[name = tensor<string, []>("var_47")];
            tensor<string, []> var_47_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_47_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1140_to_fp16 = const()[name = tensor<string, []>("op_1140_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_47_to_fp16 = cast(dtype = var_47_to_fp16_dtype_0, x = var_47)[name = tensor<string, []>("cast_581")];
            tensor<fp16, [1, 1, 1]> var_1141_cast_fp16 = add(x = var_47_to_fp16, y = var_1140_to_fp16)[name = tensor<string, []>("op_1141_cast_fp16")];
            tensor<string, []> var_1141_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1141_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1142_epsilon_0 = const()[name = tensor<string, []>("op_1142_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1141_cast_fp16_to_fp32 = cast(dtype = var_1141_cast_fp16_to_fp32_dtype_0, x = var_1141_cast_fp16)[name = tensor<string, []>("cast_580")];
            tensor<fp32, [1, 1, 1]> var_1142 = rsqrt(epsilon = var_1142_epsilon_0, x = var_1141_cast_fp16_to_fp32)[name = tensor<string, []>("op_1142")];
            tensor<string, []> var_1142_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1142_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1142_to_fp16 = cast(dtype = var_1142_to_fp16_dtype_0, x = var_1142)[name = tensor<string, []>("cast_579")];
            tensor<fp16, [1, 1, 512]> x_235_cast_fp16 = mul(x = x_229_cast_fp16, y = var_1142_to_fp16)[name = tensor<string, []>("x_235_cast_fp16")];
            tensor<fp16, [512]> layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271235520)))];
            tensor<fp16, [1, 1, 512]> x_237_cast_fp16 = mul(x = layers_5_post_attention_layernorm_weight_to_fp16, y = x_235_cast_fp16)[name = tensor<string, []>("x_237_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271236608)))];
            tensor<fp16, [1, 1, 1536]> linear_39_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_mlp_gate_proj_weight_to_fp16, x = x_237_cast_fp16)[name = tensor<string, []>("linear_39_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_1153_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor<string, []>("op_1153_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_5_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(272809536)))];
            tensor<fp16, [1, 1, 1536]> linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_mlp_up_proj_weight_to_fp16, x = x_237_cast_fp16)[name = tensor<string, []>("linear_40_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_239_cast_fp16 = mul(x = var_1153_cast_fp16, y = linear_40_cast_fp16)[name = tensor<string, []>("x_239_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_5_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_5_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(274382464)))];
            tensor<fp16, [1, 1, 512]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_5_mlp_down_proj_weight_to_fp16, x = x_239_cast_fp16)[name = tensor<string, []>("linear_41_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_241_cast_fp16 = add(x = x_229_cast_fp16, y = linear_41_cast_fp16)[name = tensor<string, []>("x_241_cast_fp16")];
            tensor<string, []> x_241_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_241_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_13_begin_0 = const()[name = tensor<string, []>("k_cache_13_begin_0"), val = tensor<int32, [5]>([6, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_13_end_0 = const()[name = tensor<string, []>("k_cache_13_end_0"), val = tensor<int32, [5]>([7, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_13_end_mask_0 = const()[name = tensor<string, []>("k_cache_13_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_13_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_13_cast_fp16 = slice_by_index(begin = k_cache_13_begin_0, end = k_cache_13_end_0, end_mask = k_cache_13_end_mask_0, squeeze_mask = k_cache_13_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_13_cast_fp16")];
            tensor<int32, [5]> v_cache_13_begin_0 = const()[name = tensor<string, []>("v_cache_13_begin_0"), val = tensor<int32, [5]>([6, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_13_end_0 = const()[name = tensor<string, []>("v_cache_13_end_0"), val = tensor<int32, [5]>([7, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_13_end_mask_0 = const()[name = tensor<string, []>("v_cache_13_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_13_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_13_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_13_cast_fp16 = slice_by_index(begin = v_cache_13_begin_0, end = v_cache_13_end_0, end_mask = v_cache_13_end_mask_0, squeeze_mask = v_cache_13_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_13_cast_fp16")];
            tensor<int32, []> var_1185 = const()[name = tensor<string, []>("op_1185"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_1184_promoted = const()[name = tensor<string, []>("op_1184_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_241_cast_fp16_to_fp32 = cast(dtype = x_241_cast_fp16_to_fp32_dtype_0, x = x_241_cast_fp16)[name = tensor<string, []>("cast_578")];
            tensor<fp32, [1, 1, 512]> var_1194 = pow(x = x_241_cast_fp16_to_fp32, y = var_1184_promoted)[name = tensor<string, []>("op_1194")];
            tensor<int32, [1]> var_49_axes_0 = const()[name = tensor<string, []>("var_49_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_49_keep_dims_0 = const()[name = tensor<string, []>("var_49_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_49 = reduce_mean(axes = var_49_axes_0, keep_dims = var_49_keep_dims_0, x = var_1194)[name = tensor<string, []>("var_49")];
            tensor<string, []> var_49_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_49_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1198_to_fp16 = const()[name = tensor<string, []>("op_1198_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_49_to_fp16 = cast(dtype = var_49_to_fp16_dtype_0, x = var_49)[name = tensor<string, []>("cast_577")];
            tensor<fp16, [1, 1, 1]> var_1199_cast_fp16 = add(x = var_49_to_fp16, y = var_1198_to_fp16)[name = tensor<string, []>("op_1199_cast_fp16")];
            tensor<string, []> var_1199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1199_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1200_epsilon_0 = const()[name = tensor<string, []>("op_1200_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1199_cast_fp16_to_fp32 = cast(dtype = var_1199_cast_fp16_to_fp32_dtype_0, x = var_1199_cast_fp16)[name = tensor<string, []>("cast_576")];
            tensor<fp32, [1, 1, 1]> var_1200 = rsqrt(epsilon = var_1200_epsilon_0, x = var_1199_cast_fp16_to_fp32)[name = tensor<string, []>("op_1200")];
            tensor<string, []> var_1200_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1200_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1200_to_fp16 = cast(dtype = var_1200_to_fp16_dtype_0, x = var_1200)[name = tensor<string, []>("cast_575")];
            tensor<fp16, [1, 1, 512]> x_247_cast_fp16 = mul(x = x_241_cast_fp16, y = var_1200_to_fp16)[name = tensor<string, []>("x_247_cast_fp16")];
            tensor<fp16, [512]> layers_6_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275955392)))];
            tensor<fp16, [1, 1, 512]> x_249_cast_fp16 = mul(x = layers_6_input_layernorm_weight_to_fp16, y = x_247_cast_fp16)[name = tensor<string, []>("x_249_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275956480)))];
            tensor<fp16, [1, 1, 1536]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = x_249_cast_fp16)[name = tensor<string, []>("linear_42_cast_fp16")];
            tensor<int32, [4]> var_1214 = const()[name = tensor<string, []>("op_1214"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_251_cast_fp16 = reshape(shape = var_1214, x = linear_42_cast_fp16)[name = tensor<string, []>("x_251_cast_fp16")];
            tensor<string, []> x_251_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_251_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1184_promoted_1 = const()[name = tensor<string, []>("op_1184_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_251_cast_fp16_to_fp32 = cast(dtype = x_251_cast_fp16_to_fp32_dtype_0, x = x_251_cast_fp16)[name = tensor<string, []>("cast_574")];
            tensor<fp32, [1, 1, 12, 128]> var_1218 = pow(x = x_251_cast_fp16_to_fp32, y = var_1184_promoted_1)[name = tensor<string, []>("op_1218")];
            tensor<int32, [1]> var_51_axes_0 = const()[name = tensor<string, []>("var_51_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_51_keep_dims_0 = const()[name = tensor<string, []>("var_51_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_51 = reduce_mean(axes = var_51_axes_0, keep_dims = var_51_keep_dims_0, x = var_1218)[name = tensor<string, []>("var_51")];
            tensor<string, []> var_51_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_51_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1222_to_fp16 = const()[name = tensor<string, []>("op_1222_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_51_to_fp16 = cast(dtype = var_51_to_fp16_dtype_0, x = var_51)[name = tensor<string, []>("cast_573")];
            tensor<fp16, [1, 1, 12, 1]> var_1223_cast_fp16 = add(x = var_51_to_fp16, y = var_1222_to_fp16)[name = tensor<string, []>("op_1223_cast_fp16")];
            tensor<string, []> var_1223_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1223_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1224_epsilon_0 = const()[name = tensor<string, []>("op_1224_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_1223_cast_fp16_to_fp32 = cast(dtype = var_1223_cast_fp16_to_fp32_dtype_0, x = var_1223_cast_fp16)[name = tensor<string, []>("cast_572")];
            tensor<fp32, [1, 1, 12, 1]> var_1224 = rsqrt(epsilon = var_1224_epsilon_0, x = var_1223_cast_fp16_to_fp32)[name = tensor<string, []>("op_1224")];
            tensor<string, []> var_1224_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1224_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_1224_to_fp16 = cast(dtype = var_1224_to_fp16_dtype_0, x = var_1224)[name = tensor<string, []>("cast_571")];
            tensor<fp16, [1, 1, 12, 128]> x_257_cast_fp16 = mul(x = x_251_cast_fp16, y = var_1224_to_fp16)[name = tensor<string, []>("x_257_cast_fp16")];
            tensor<fp16, [128]> layers_6_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277529408)))];
            tensor<fp16, [1, 1, 12, 128]> var_1226_cast_fp16 = mul(x = layers_6_self_attn_q_norm_weight_to_fp16, y = x_257_cast_fp16)[name = tensor<string, []>("op_1226_cast_fp16")];
            tensor<int32, [4]> q_25_perm_0 = const()[name = tensor<string, []>("q_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(277529728)))];
            tensor<fp16, [1, 1, 512]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = x_249_cast_fp16)[name = tensor<string, []>("linear_43_cast_fp16")];
            tensor<int32, [4]> var_1230 = const()[name = tensor<string, []>("op_1230"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_259_cast_fp16 = reshape(shape = var_1230, x = linear_43_cast_fp16)[name = tensor<string, []>("x_259_cast_fp16")];
            tensor<string, []> x_259_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_259_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1184_promoted_2 = const()[name = tensor<string, []>("op_1184_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_259_cast_fp16_to_fp32 = cast(dtype = x_259_cast_fp16_to_fp32_dtype_0, x = x_259_cast_fp16)[name = tensor<string, []>("cast_570")];
            tensor<fp32, [1, 1, 4, 128]> var_1234 = pow(x = x_259_cast_fp16_to_fp32, y = var_1184_promoted_2)[name = tensor<string, []>("op_1234")];
            tensor<int32, [1]> var_53_axes_0 = const()[name = tensor<string, []>("var_53_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_53_keep_dims_0 = const()[name = tensor<string, []>("var_53_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_53 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = var_1234)[name = tensor<string, []>("var_53")];
            tensor<string, []> var_53_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_53_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1238_to_fp16 = const()[name = tensor<string, []>("op_1238_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_53_to_fp16 = cast(dtype = var_53_to_fp16_dtype_0, x = var_53)[name = tensor<string, []>("cast_569")];
            tensor<fp16, [1, 1, 4, 1]> var_1239_cast_fp16 = add(x = var_53_to_fp16, y = var_1238_to_fp16)[name = tensor<string, []>("op_1239_cast_fp16")];
            tensor<string, []> var_1239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1239_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1240_epsilon_0 = const()[name = tensor<string, []>("op_1240_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_1239_cast_fp16_to_fp32 = cast(dtype = var_1239_cast_fp16_to_fp32_dtype_0, x = var_1239_cast_fp16)[name = tensor<string, []>("cast_568")];
            tensor<fp32, [1, 1, 4, 1]> var_1240 = rsqrt(epsilon = var_1240_epsilon_0, x = var_1239_cast_fp16_to_fp32)[name = tensor<string, []>("op_1240")];
            tensor<string, []> var_1240_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1240_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_1240_to_fp16 = cast(dtype = var_1240_to_fp16_dtype_0, x = var_1240)[name = tensor<string, []>("cast_567")];
            tensor<fp16, [1, 1, 4, 128]> x_265_cast_fp16 = mul(x = x_259_cast_fp16, y = var_1240_to_fp16)[name = tensor<string, []>("x_265_cast_fp16")];
            tensor<fp16, [128]> layers_6_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278054080)))];
            tensor<fp16, [1, 1, 4, 128]> var_1242_cast_fp16 = mul(x = layers_6_self_attn_k_norm_weight_to_fp16, y = x_265_cast_fp16)[name = tensor<string, []>("op_1242_cast_fp16")];
            tensor<int32, [4]> k_25_perm_0 = const()[name = tensor<string, []>("k_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278054400)))];
            tensor<fp16, [1, 1, 512]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = x_249_cast_fp16)[name = tensor<string, []>("linear_44_cast_fp16")];
            tensor<int32, [4]> var_1246 = const()[name = tensor<string, []>("op_1246"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_1247_cast_fp16 = reshape(shape = var_1246, x = linear_44_cast_fp16)[name = tensor<string, []>("op_1247_cast_fp16")];
            tensor<int32, [4]> v_13_perm_0 = const()[name = tensor<string, []>("v_13_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_25_cast_fp16 = transpose(perm = q_25_perm_0, x = var_1226_cast_fp16)[name = tensor<string, []>("transpose_87")];
            tensor<fp16, [1, 12, 1, 128]> var_1251_cast_fp16 = mul(x = q_25_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1251_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = tensor<string, []>("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = tensor<string, []>("x1_25_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = tensor<string, []>("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_25_cast_fp16)[name = tensor<string, []>("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = tensor<string, []>("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = tensor<string, []>("x2_25_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = tensor<string, []>("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_25_cast_fp16)[name = tensor<string, []>("x2_25_cast_fp16")];
            tensor<fp16, []> const_13_promoted_to_fp16 = const()[name = tensor<string, []>("const_13_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_1254_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_13_promoted_to_fp16)[name = tensor<string, []>("op_1254_cast_fp16")];
            tensor<bool, []> var_1256_interleave_0 = const()[name = tensor<string, []>("op_1256_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_1256_cast_fp16 = concat(axis = var_1185, interleave = var_1256_interleave_0, values = (var_1254_cast_fp16, x1_25_cast_fp16))[name = tensor<string, []>("op_1256_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_1257_cast_fp16 = mul(x = var_1256_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1257_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_27_cast_fp16 = add(x = var_1251_cast_fp16, y = var_1257_cast_fp16)[name = tensor<string, []>("q_27_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_25_cast_fp16 = transpose(perm = k_25_perm_0, x = var_1242_cast_fp16)[name = tensor<string, []>("transpose_86")];
            tensor<fp16, [1, 4, 1, 128]> var_1259_cast_fp16 = mul(x = k_25_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1259_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = tensor<string, []>("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = tensor<string, []>("x1_27_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = tensor<string, []>("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_25_cast_fp16)[name = tensor<string, []>("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = tensor<string, []>("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = tensor<string, []>("x2_27_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = tensor<string, []>("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_25_cast_fp16)[name = tensor<string, []>("x2_27_cast_fp16")];
            tensor<fp16, []> const_14_promoted_to_fp16 = const()[name = tensor<string, []>("const_14_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_1262_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_14_promoted_to_fp16)[name = tensor<string, []>("op_1262_cast_fp16")];
            tensor<bool, []> var_1264_interleave_0 = const()[name = tensor<string, []>("op_1264_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_1264_cast_fp16 = concat(axis = var_1185, interleave = var_1264_interleave_0, values = (var_1262_cast_fp16, x1_27_cast_fp16))[name = tensor<string, []>("op_1264_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_1265_cast_fp16 = mul(x = var_1264_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1265_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_27_cast_fp16 = add(x = var_1259_cast_fp16, y = var_1265_cast_fp16)[name = tensor<string, []>("k_27_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1268_cast_fp16 = mul(x = k_cache_13_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1268_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1269_cast_fp16 = mul(x = k_27_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1269_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_13_cast_fp16 = add(x = var_1268_cast_fp16, y = var_1269_cast_fp16)[name = tensor<string, []>("k_full_13_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1272_cast_fp16 = mul(x = v_cache_13_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1272_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_1247_cast_fp16)[name = tensor<string, []>("transpose_85")];
            tensor<fp16, [1, 4, 2048, 128]> var_1273_cast_fp16 = mul(x = v_13_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1273_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_13_cast_fp16 = add(x = var_1272_cast_fp16, y = var_1273_cast_fp16)[name = tensor<string, []>("v_full_13_cast_fp16")];
            tensor<int32, [1]> var_1275_axes_0 = const()[name = tensor<string, []>("op_1275_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1275_cast_fp16 = expand_dims(axes = var_1275_axes_0, x = k_full_13_cast_fp16)[name = tensor<string, []>("op_1275_cast_fp16")];
            tensor<int32, [5]> var_1277_reps_0 = const()[name = tensor<string, []>("op_1277_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1277_cast_fp16 = tile(reps = var_1277_reps_0, x = var_1275_cast_fp16)[name = tensor<string, []>("op_1277_cast_fp16")];
            tensor<int32, [4]> var_1278 = const()[name = tensor<string, []>("op_1278"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_13_cast_fp16 = reshape(shape = var_1278, x = var_1277_cast_fp16)[name = tensor<string, []>("k_rep_13_cast_fp16")];
            tensor<int32, [1]> var_1280_axes_0 = const()[name = tensor<string, []>("op_1280_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1280_cast_fp16 = expand_dims(axes = var_1280_axes_0, x = v_full_13_cast_fp16)[name = tensor<string, []>("op_1280_cast_fp16")];
            tensor<int32, [5]> var_1282_reps_0 = const()[name = tensor<string, []>("op_1282_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1282_cast_fp16 = tile(reps = var_1282_reps_0, x = var_1280_cast_fp16)[name = tensor<string, []>("op_1282_cast_fp16")];
            tensor<int32, [4]> var_1283 = const()[name = tensor<string, []>("op_1283"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_13_cast_fp16 = reshape(shape = var_1283, x = var_1282_cast_fp16)[name = tensor<string, []>("v_rep_13_cast_fp16")];
            tensor<bool, []> var_1286_transpose_x_1 = const()[name = tensor<string, []>("op_1286_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1286_transpose_y_1 = const()[name = tensor<string, []>("op_1286_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_1286_cast_fp16 = matmul(transpose_x = var_1286_transpose_x_1, transpose_y = var_1286_transpose_y_1, x = q_27_cast_fp16, y = k_rep_13_cast_fp16)[name = tensor<string, []>("op_1286_cast_fp16")];
            tensor<fp16, []> var_1287_to_fp16 = const()[name = tensor<string, []>("op_1287_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_25_cast_fp16 = mul(x = var_1286_cast_fp16, y = var_1287_to_fp16)[name = tensor<string, []>("attn_25_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_27_cast_fp16 = add(x = attn_25_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_27_cast_fp16")];
            tensor<string, []> input_27_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_27_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_27_cast_fp16_to_fp32 = cast(dtype = input_27_cast_fp16_to_fp32_dtype_0, x = input_27_cast_fp16)[name = tensor<string, []>("cast_566")];
            tensor<fp32, [1, 12, 1, 2048]> attn_27 = softmax(axis = var_1185, x = input_27_cast_fp16_to_fp32)[name = tensor<string, []>("attn_27")];
            tensor<bool, []> out_13_transpose_x_0 = const()[name = tensor<string, []>("out_13_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_13_transpose_y_0 = const()[name = tensor<string, []>("out_13_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_27_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_27_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_27_to_fp16 = cast(dtype = attn_27_to_fp16_dtype_0, x = attn_27)[name = tensor<string, []>("cast_565")];
            tensor<fp16, [1, 12, 1, 128]> out_13_cast_fp16 = matmul(transpose_x = out_13_transpose_x_0, transpose_y = out_13_transpose_y_0, x = attn_27_to_fp16, y = v_rep_13_cast_fp16)[name = tensor<string, []>("out_13_cast_fp16")];
            tensor<int32, [4]> var_1292_perm_0 = const()[name = tensor<string, []>("op_1292_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1294 = const()[name = tensor<string, []>("op_1294"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_1292_cast_fp16 = transpose(perm = var_1292_perm_0, x = out_13_cast_fp16)[name = tensor<string, []>("transpose_84")];
            tensor<fp16, [1, 1, 1536]> x_267_cast_fp16 = reshape(shape = var_1294, x = var_1292_cast_fp16)[name = tensor<string, []>("x_267_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(278578752)))];
            tensor<fp16, [1, 1, 512]> linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = x_267_cast_fp16)[name = tensor<string, []>("linear_45_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_269_cast_fp16 = add(x = x_241_cast_fp16, y = linear_45_cast_fp16)[name = tensor<string, []>("x_269_cast_fp16")];
            tensor<string, []> x_269_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_269_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1184_promoted_3 = const()[name = tensor<string, []>("op_1184_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_269_cast_fp16_to_fp32 = cast(dtype = x_269_cast_fp16_to_fp32_dtype_0, x = x_269_cast_fp16)[name = tensor<string, []>("cast_564")];
            tensor<fp32, [1, 1, 512]> var_1305 = pow(x = x_269_cast_fp16_to_fp32, y = var_1184_promoted_3)[name = tensor<string, []>("op_1305")];
            tensor<int32, [1]> var_55_axes_0 = const()[name = tensor<string, []>("var_55_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_55_keep_dims_0 = const()[name = tensor<string, []>("var_55_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_55 = reduce_mean(axes = var_55_axes_0, keep_dims = var_55_keep_dims_0, x = var_1305)[name = tensor<string, []>("var_55")];
            tensor<string, []> var_55_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_55_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1309_to_fp16 = const()[name = tensor<string, []>("op_1309_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_55_to_fp16 = cast(dtype = var_55_to_fp16_dtype_0, x = var_55)[name = tensor<string, []>("cast_563")];
            tensor<fp16, [1, 1, 1]> var_1310_cast_fp16 = add(x = var_55_to_fp16, y = var_1309_to_fp16)[name = tensor<string, []>("op_1310_cast_fp16")];
            tensor<string, []> var_1310_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1310_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1311_epsilon_0 = const()[name = tensor<string, []>("op_1311_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1310_cast_fp16_to_fp32 = cast(dtype = var_1310_cast_fp16_to_fp32_dtype_0, x = var_1310_cast_fp16)[name = tensor<string, []>("cast_562")];
            tensor<fp32, [1, 1, 1]> var_1311 = rsqrt(epsilon = var_1311_epsilon_0, x = var_1310_cast_fp16_to_fp32)[name = tensor<string, []>("op_1311")];
            tensor<string, []> var_1311_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1311_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1311_to_fp16 = cast(dtype = var_1311_to_fp16_dtype_0, x = var_1311)[name = tensor<string, []>("cast_561")];
            tensor<fp16, [1, 1, 512]> x_275_cast_fp16 = mul(x = x_269_cast_fp16, y = var_1311_to_fp16)[name = tensor<string, []>("x_275_cast_fp16")];
            tensor<fp16, [512]> layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280151680)))];
            tensor<fp16, [1, 1, 512]> x_277_cast_fp16 = mul(x = layers_6_post_attention_layernorm_weight_to_fp16, y = x_275_cast_fp16)[name = tensor<string, []>("x_277_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_6_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280152768)))];
            tensor<fp16, [1, 1, 1536]> linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_mlp_gate_proj_weight_to_fp16, x = x_277_cast_fp16)[name = tensor<string, []>("linear_46_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_1322_cast_fp16 = silu(x = linear_46_cast_fp16)[name = tensor<string, []>("op_1322_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_6_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281725696)))];
            tensor<fp16, [1, 1, 1536]> linear_47_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_mlp_up_proj_weight_to_fp16, x = x_277_cast_fp16)[name = tensor<string, []>("linear_47_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_279_cast_fp16 = mul(x = var_1322_cast_fp16, y = linear_47_cast_fp16)[name = tensor<string, []>("x_279_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_6_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_6_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283298624)))];
            tensor<fp16, [1, 1, 512]> linear_48_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_6_mlp_down_proj_weight_to_fp16, x = x_279_cast_fp16)[name = tensor<string, []>("linear_48_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_281_cast_fp16 = add(x = x_269_cast_fp16, y = linear_48_cast_fp16)[name = tensor<string, []>("x_281_cast_fp16")];
            tensor<string, []> x_281_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_281_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_15_begin_0 = const()[name = tensor<string, []>("k_cache_15_begin_0"), val = tensor<int32, [5]>([7, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_15_end_0 = const()[name = tensor<string, []>("k_cache_15_end_0"), val = tensor<int32, [5]>([8, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_15_end_mask_0 = const()[name = tensor<string, []>("k_cache_15_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_15_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_15_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_15_cast_fp16 = slice_by_index(begin = k_cache_15_begin_0, end = k_cache_15_end_0, end_mask = k_cache_15_end_mask_0, squeeze_mask = k_cache_15_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_15_cast_fp16")];
            tensor<int32, [5]> v_cache_15_begin_0 = const()[name = tensor<string, []>("v_cache_15_begin_0"), val = tensor<int32, [5]>([7, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_15_end_0 = const()[name = tensor<string, []>("v_cache_15_end_0"), val = tensor<int32, [5]>([8, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_15_end_mask_0 = const()[name = tensor<string, []>("v_cache_15_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_15_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_15_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_15_cast_fp16 = slice_by_index(begin = v_cache_15_begin_0, end = v_cache_15_end_0, end_mask = v_cache_15_end_mask_0, squeeze_mask = v_cache_15_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_15_cast_fp16")];
            tensor<int32, []> var_1354 = const()[name = tensor<string, []>("op_1354"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_1353_promoted = const()[name = tensor<string, []>("op_1353_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_281_cast_fp16_to_fp32 = cast(dtype = x_281_cast_fp16_to_fp32_dtype_0, x = x_281_cast_fp16)[name = tensor<string, []>("cast_560")];
            tensor<fp32, [1, 1, 512]> var_1363 = pow(x = x_281_cast_fp16_to_fp32, y = var_1353_promoted)[name = tensor<string, []>("op_1363")];
            tensor<int32, [1]> var_57_axes_0 = const()[name = tensor<string, []>("var_57_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_57_keep_dims_0 = const()[name = tensor<string, []>("var_57_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_57 = reduce_mean(axes = var_57_axes_0, keep_dims = var_57_keep_dims_0, x = var_1363)[name = tensor<string, []>("var_57")];
            tensor<string, []> var_57_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_57_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1367_to_fp16 = const()[name = tensor<string, []>("op_1367_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_57_to_fp16 = cast(dtype = var_57_to_fp16_dtype_0, x = var_57)[name = tensor<string, []>("cast_559")];
            tensor<fp16, [1, 1, 1]> var_1368_cast_fp16 = add(x = var_57_to_fp16, y = var_1367_to_fp16)[name = tensor<string, []>("op_1368_cast_fp16")];
            tensor<string, []> var_1368_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1368_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1369_epsilon_0 = const()[name = tensor<string, []>("op_1369_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1368_cast_fp16_to_fp32 = cast(dtype = var_1368_cast_fp16_to_fp32_dtype_0, x = var_1368_cast_fp16)[name = tensor<string, []>("cast_558")];
            tensor<fp32, [1, 1, 1]> var_1369 = rsqrt(epsilon = var_1369_epsilon_0, x = var_1368_cast_fp16_to_fp32)[name = tensor<string, []>("op_1369")];
            tensor<string, []> var_1369_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1369_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1369_to_fp16 = cast(dtype = var_1369_to_fp16_dtype_0, x = var_1369)[name = tensor<string, []>("cast_557")];
            tensor<fp16, [1, 1, 512]> x_287_cast_fp16 = mul(x = x_281_cast_fp16, y = var_1369_to_fp16)[name = tensor<string, []>("x_287_cast_fp16")];
            tensor<fp16, [512]> layers_7_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284871552)))];
            tensor<fp16, [1, 1, 512]> x_289_cast_fp16 = mul(x = layers_7_input_layernorm_weight_to_fp16, y = x_287_cast_fp16)[name = tensor<string, []>("x_289_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284872640)))];
            tensor<fp16, [1, 1, 1536]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = x_289_cast_fp16)[name = tensor<string, []>("linear_49_cast_fp16")];
            tensor<int32, [4]> var_1383 = const()[name = tensor<string, []>("op_1383"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_291_cast_fp16 = reshape(shape = var_1383, x = linear_49_cast_fp16)[name = tensor<string, []>("x_291_cast_fp16")];
            tensor<string, []> x_291_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_291_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1353_promoted_1 = const()[name = tensor<string, []>("op_1353_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_291_cast_fp16_to_fp32 = cast(dtype = x_291_cast_fp16_to_fp32_dtype_0, x = x_291_cast_fp16)[name = tensor<string, []>("cast_556")];
            tensor<fp32, [1, 1, 12, 128]> var_1387 = pow(x = x_291_cast_fp16_to_fp32, y = var_1353_promoted_1)[name = tensor<string, []>("op_1387")];
            tensor<int32, [1]> var_59_axes_0 = const()[name = tensor<string, []>("var_59_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_59_keep_dims_0 = const()[name = tensor<string, []>("var_59_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_59 = reduce_mean(axes = var_59_axes_0, keep_dims = var_59_keep_dims_0, x = var_1387)[name = tensor<string, []>("var_59")];
            tensor<string, []> var_59_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_59_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1391_to_fp16 = const()[name = tensor<string, []>("op_1391_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_59_to_fp16 = cast(dtype = var_59_to_fp16_dtype_0, x = var_59)[name = tensor<string, []>("cast_555")];
            tensor<fp16, [1, 1, 12, 1]> var_1392_cast_fp16 = add(x = var_59_to_fp16, y = var_1391_to_fp16)[name = tensor<string, []>("op_1392_cast_fp16")];
            tensor<string, []> var_1392_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1392_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1393_epsilon_0 = const()[name = tensor<string, []>("op_1393_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_1392_cast_fp16_to_fp32 = cast(dtype = var_1392_cast_fp16_to_fp32_dtype_0, x = var_1392_cast_fp16)[name = tensor<string, []>("cast_554")];
            tensor<fp32, [1, 1, 12, 1]> var_1393 = rsqrt(epsilon = var_1393_epsilon_0, x = var_1392_cast_fp16_to_fp32)[name = tensor<string, []>("op_1393")];
            tensor<string, []> var_1393_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1393_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_1393_to_fp16 = cast(dtype = var_1393_to_fp16_dtype_0, x = var_1393)[name = tensor<string, []>("cast_553")];
            tensor<fp16, [1, 1, 12, 128]> x_297_cast_fp16 = mul(x = x_291_cast_fp16, y = var_1393_to_fp16)[name = tensor<string, []>("x_297_cast_fp16")];
            tensor<fp16, [128]> layers_7_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286445568)))];
            tensor<fp16, [1, 1, 12, 128]> var_1395_cast_fp16 = mul(x = layers_7_self_attn_q_norm_weight_to_fp16, y = x_297_cast_fp16)[name = tensor<string, []>("op_1395_cast_fp16")];
            tensor<int32, [4]> q_29_perm_0 = const()[name = tensor<string, []>("q_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286445888)))];
            tensor<fp16, [1, 1, 512]> linear_50_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = x_289_cast_fp16)[name = tensor<string, []>("linear_50_cast_fp16")];
            tensor<int32, [4]> var_1399 = const()[name = tensor<string, []>("op_1399"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_299_cast_fp16 = reshape(shape = var_1399, x = linear_50_cast_fp16)[name = tensor<string, []>("x_299_cast_fp16")];
            tensor<string, []> x_299_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_299_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1353_promoted_2 = const()[name = tensor<string, []>("op_1353_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_299_cast_fp16_to_fp32 = cast(dtype = x_299_cast_fp16_to_fp32_dtype_0, x = x_299_cast_fp16)[name = tensor<string, []>("cast_552")];
            tensor<fp32, [1, 1, 4, 128]> var_1403 = pow(x = x_299_cast_fp16_to_fp32, y = var_1353_promoted_2)[name = tensor<string, []>("op_1403")];
            tensor<int32, [1]> var_61_axes_0 = const()[name = tensor<string, []>("var_61_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_61_keep_dims_0 = const()[name = tensor<string, []>("var_61_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_61 = reduce_mean(axes = var_61_axes_0, keep_dims = var_61_keep_dims_0, x = var_1403)[name = tensor<string, []>("var_61")];
            tensor<string, []> var_61_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_61_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1407_to_fp16 = const()[name = tensor<string, []>("op_1407_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_61_to_fp16 = cast(dtype = var_61_to_fp16_dtype_0, x = var_61)[name = tensor<string, []>("cast_551")];
            tensor<fp16, [1, 1, 4, 1]> var_1408_cast_fp16 = add(x = var_61_to_fp16, y = var_1407_to_fp16)[name = tensor<string, []>("op_1408_cast_fp16")];
            tensor<string, []> var_1408_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1408_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1409_epsilon_0 = const()[name = tensor<string, []>("op_1409_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_1408_cast_fp16_to_fp32 = cast(dtype = var_1408_cast_fp16_to_fp32_dtype_0, x = var_1408_cast_fp16)[name = tensor<string, []>("cast_550")];
            tensor<fp32, [1, 1, 4, 1]> var_1409 = rsqrt(epsilon = var_1409_epsilon_0, x = var_1408_cast_fp16_to_fp32)[name = tensor<string, []>("op_1409")];
            tensor<string, []> var_1409_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1409_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_1409_to_fp16 = cast(dtype = var_1409_to_fp16_dtype_0, x = var_1409)[name = tensor<string, []>("cast_549")];
            tensor<fp16, [1, 1, 4, 128]> x_305_cast_fp16 = mul(x = x_299_cast_fp16, y = var_1409_to_fp16)[name = tensor<string, []>("x_305_cast_fp16")];
            tensor<fp16, [128]> layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286970240)))];
            tensor<fp16, [1, 1, 4, 128]> var_1411_cast_fp16 = mul(x = layers_7_self_attn_k_norm_weight_to_fp16, y = x_305_cast_fp16)[name = tensor<string, []>("op_1411_cast_fp16")];
            tensor<int32, [4]> k_29_perm_0 = const()[name = tensor<string, []>("k_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286970560)))];
            tensor<fp16, [1, 1, 512]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = x_289_cast_fp16)[name = tensor<string, []>("linear_51_cast_fp16")];
            tensor<int32, [4]> var_1415 = const()[name = tensor<string, []>("op_1415"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_1416_cast_fp16 = reshape(shape = var_1415, x = linear_51_cast_fp16)[name = tensor<string, []>("op_1416_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = tensor<string, []>("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_29_cast_fp16 = transpose(perm = q_29_perm_0, x = var_1395_cast_fp16)[name = tensor<string, []>("transpose_83")];
            tensor<fp16, [1, 12, 1, 128]> var_1420_cast_fp16 = mul(x = q_29_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1420_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = tensor<string, []>("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = tensor<string, []>("x1_29_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = tensor<string, []>("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_29_cast_fp16)[name = tensor<string, []>("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = tensor<string, []>("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = tensor<string, []>("x2_29_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = tensor<string, []>("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_29_cast_fp16)[name = tensor<string, []>("x2_29_cast_fp16")];
            tensor<fp16, []> const_15_promoted_to_fp16 = const()[name = tensor<string, []>("const_15_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_1423_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_15_promoted_to_fp16)[name = tensor<string, []>("op_1423_cast_fp16")];
            tensor<bool, []> var_1425_interleave_0 = const()[name = tensor<string, []>("op_1425_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_1425_cast_fp16 = concat(axis = var_1354, interleave = var_1425_interleave_0, values = (var_1423_cast_fp16, x1_29_cast_fp16))[name = tensor<string, []>("op_1425_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_1426_cast_fp16 = mul(x = var_1425_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1426_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_31_cast_fp16 = add(x = var_1420_cast_fp16, y = var_1426_cast_fp16)[name = tensor<string, []>("q_31_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_29_cast_fp16 = transpose(perm = k_29_perm_0, x = var_1411_cast_fp16)[name = tensor<string, []>("transpose_82")];
            tensor<fp16, [1, 4, 1, 128]> var_1428_cast_fp16 = mul(x = k_29_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1428_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = tensor<string, []>("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = tensor<string, []>("x1_31_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = tensor<string, []>("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_29_cast_fp16)[name = tensor<string, []>("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = tensor<string, []>("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = tensor<string, []>("x2_31_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = tensor<string, []>("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_29_cast_fp16)[name = tensor<string, []>("x2_31_cast_fp16")];
            tensor<fp16, []> const_16_promoted_to_fp16 = const()[name = tensor<string, []>("const_16_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_1431_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_16_promoted_to_fp16)[name = tensor<string, []>("op_1431_cast_fp16")];
            tensor<bool, []> var_1433_interleave_0 = const()[name = tensor<string, []>("op_1433_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_1433_cast_fp16 = concat(axis = var_1354, interleave = var_1433_interleave_0, values = (var_1431_cast_fp16, x1_31_cast_fp16))[name = tensor<string, []>("op_1433_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_1434_cast_fp16 = mul(x = var_1433_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1434_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_31_cast_fp16 = add(x = var_1428_cast_fp16, y = var_1434_cast_fp16)[name = tensor<string, []>("k_31_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1437_cast_fp16 = mul(x = k_cache_15_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1437_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1438_cast_fp16 = mul(x = k_31_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1438_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_15_cast_fp16 = add(x = var_1437_cast_fp16, y = var_1438_cast_fp16)[name = tensor<string, []>("k_full_15_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1441_cast_fp16 = mul(x = v_cache_15_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1441_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_1416_cast_fp16)[name = tensor<string, []>("transpose_81")];
            tensor<fp16, [1, 4, 2048, 128]> var_1442_cast_fp16 = mul(x = v_15_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1442_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_15_cast_fp16 = add(x = var_1441_cast_fp16, y = var_1442_cast_fp16)[name = tensor<string, []>("v_full_15_cast_fp16")];
            tensor<int32, [1]> var_1444_axes_0 = const()[name = tensor<string, []>("op_1444_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1444_cast_fp16 = expand_dims(axes = var_1444_axes_0, x = k_full_15_cast_fp16)[name = tensor<string, []>("op_1444_cast_fp16")];
            tensor<int32, [5]> var_1446_reps_0 = const()[name = tensor<string, []>("op_1446_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1446_cast_fp16 = tile(reps = var_1446_reps_0, x = var_1444_cast_fp16)[name = tensor<string, []>("op_1446_cast_fp16")];
            tensor<int32, [4]> var_1447 = const()[name = tensor<string, []>("op_1447"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_15_cast_fp16 = reshape(shape = var_1447, x = var_1446_cast_fp16)[name = tensor<string, []>("k_rep_15_cast_fp16")];
            tensor<int32, [1]> var_1449_axes_0 = const()[name = tensor<string, []>("op_1449_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1449_cast_fp16 = expand_dims(axes = var_1449_axes_0, x = v_full_15_cast_fp16)[name = tensor<string, []>("op_1449_cast_fp16")];
            tensor<int32, [5]> var_1451_reps_0 = const()[name = tensor<string, []>("op_1451_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1451_cast_fp16 = tile(reps = var_1451_reps_0, x = var_1449_cast_fp16)[name = tensor<string, []>("op_1451_cast_fp16")];
            tensor<int32, [4]> var_1452 = const()[name = tensor<string, []>("op_1452"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_15_cast_fp16 = reshape(shape = var_1452, x = var_1451_cast_fp16)[name = tensor<string, []>("v_rep_15_cast_fp16")];
            tensor<bool, []> var_1455_transpose_x_1 = const()[name = tensor<string, []>("op_1455_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1455_transpose_y_1 = const()[name = tensor<string, []>("op_1455_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_1455_cast_fp16 = matmul(transpose_x = var_1455_transpose_x_1, transpose_y = var_1455_transpose_y_1, x = q_31_cast_fp16, y = k_rep_15_cast_fp16)[name = tensor<string, []>("op_1455_cast_fp16")];
            tensor<fp16, []> var_1456_to_fp16 = const()[name = tensor<string, []>("op_1456_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_29_cast_fp16 = mul(x = var_1455_cast_fp16, y = var_1456_to_fp16)[name = tensor<string, []>("attn_29_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_31_cast_fp16 = add(x = attn_29_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_31_cast_fp16")];
            tensor<string, []> input_31_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_31_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_31_cast_fp16_to_fp32 = cast(dtype = input_31_cast_fp16_to_fp32_dtype_0, x = input_31_cast_fp16)[name = tensor<string, []>("cast_548")];
            tensor<fp32, [1, 12, 1, 2048]> attn_31 = softmax(axis = var_1354, x = input_31_cast_fp16_to_fp32)[name = tensor<string, []>("attn_31")];
            tensor<bool, []> out_15_transpose_x_0 = const()[name = tensor<string, []>("out_15_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_15_transpose_y_0 = const()[name = tensor<string, []>("out_15_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_31_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_31_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_31_to_fp16 = cast(dtype = attn_31_to_fp16_dtype_0, x = attn_31)[name = tensor<string, []>("cast_547")];
            tensor<fp16, [1, 12, 1, 128]> out_15_cast_fp16 = matmul(transpose_x = out_15_transpose_x_0, transpose_y = out_15_transpose_y_0, x = attn_31_to_fp16, y = v_rep_15_cast_fp16)[name = tensor<string, []>("out_15_cast_fp16")];
            tensor<int32, [4]> var_1461_perm_0 = const()[name = tensor<string, []>("op_1461_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1463 = const()[name = tensor<string, []>("op_1463"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_1461_cast_fp16 = transpose(perm = var_1461_perm_0, x = out_15_cast_fp16)[name = tensor<string, []>("transpose_80")];
            tensor<fp16, [1, 1, 1536]> x_307_cast_fp16 = reshape(shape = var_1463, x = var_1461_cast_fp16)[name = tensor<string, []>("x_307_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287494912)))];
            tensor<fp16, [1, 1, 512]> linear_52_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = x_307_cast_fp16)[name = tensor<string, []>("linear_52_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_309_cast_fp16 = add(x = x_281_cast_fp16, y = linear_52_cast_fp16)[name = tensor<string, []>("x_309_cast_fp16")];
            tensor<string, []> x_309_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_309_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1353_promoted_3 = const()[name = tensor<string, []>("op_1353_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_309_cast_fp16_to_fp32 = cast(dtype = x_309_cast_fp16_to_fp32_dtype_0, x = x_309_cast_fp16)[name = tensor<string, []>("cast_546")];
            tensor<fp32, [1, 1, 512]> var_1474 = pow(x = x_309_cast_fp16_to_fp32, y = var_1353_promoted_3)[name = tensor<string, []>("op_1474")];
            tensor<int32, [1]> var_63_axes_0 = const()[name = tensor<string, []>("var_63_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_63_keep_dims_0 = const()[name = tensor<string, []>("var_63_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_63 = reduce_mean(axes = var_63_axes_0, keep_dims = var_63_keep_dims_0, x = var_1474)[name = tensor<string, []>("var_63")];
            tensor<string, []> var_63_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_63_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1478_to_fp16 = const()[name = tensor<string, []>("op_1478_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_63_to_fp16 = cast(dtype = var_63_to_fp16_dtype_0, x = var_63)[name = tensor<string, []>("cast_545")];
            tensor<fp16, [1, 1, 1]> var_1479_cast_fp16 = add(x = var_63_to_fp16, y = var_1478_to_fp16)[name = tensor<string, []>("op_1479_cast_fp16")];
            tensor<string, []> var_1479_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1479_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1480_epsilon_0 = const()[name = tensor<string, []>("op_1480_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1479_cast_fp16_to_fp32 = cast(dtype = var_1479_cast_fp16_to_fp32_dtype_0, x = var_1479_cast_fp16)[name = tensor<string, []>("cast_544")];
            tensor<fp32, [1, 1, 1]> var_1480 = rsqrt(epsilon = var_1480_epsilon_0, x = var_1479_cast_fp16_to_fp32)[name = tensor<string, []>("op_1480")];
            tensor<string, []> var_1480_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1480_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1480_to_fp16 = cast(dtype = var_1480_to_fp16_dtype_0, x = var_1480)[name = tensor<string, []>("cast_543")];
            tensor<fp16, [1, 1, 512]> x_315_cast_fp16 = mul(x = x_309_cast_fp16, y = var_1480_to_fp16)[name = tensor<string, []>("x_315_cast_fp16")];
            tensor<fp16, [512]> layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289067840)))];
            tensor<fp16, [1, 1, 512]> x_317_cast_fp16 = mul(x = layers_7_post_attention_layernorm_weight_to_fp16, y = x_315_cast_fp16)[name = tensor<string, []>("x_317_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_7_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(289068928)))];
            tensor<fp16, [1, 1, 1536]> linear_53_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_mlp_gate_proj_weight_to_fp16, x = x_317_cast_fp16)[name = tensor<string, []>("linear_53_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_1491_cast_fp16 = silu(x = linear_53_cast_fp16)[name = tensor<string, []>("op_1491_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(290641856)))];
            tensor<fp16, [1, 1, 1536]> linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_mlp_up_proj_weight_to_fp16, x = x_317_cast_fp16)[name = tensor<string, []>("linear_54_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_319_cast_fp16 = mul(x = var_1491_cast_fp16, y = linear_54_cast_fp16)[name = tensor<string, []>("x_319_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(292214784)))];
            tensor<fp16, [1, 1, 512]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_7_mlp_down_proj_weight_to_fp16, x = x_319_cast_fp16)[name = tensor<string, []>("linear_55_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_321_cast_fp16 = add(x = x_309_cast_fp16, y = linear_55_cast_fp16)[name = tensor<string, []>("x_321_cast_fp16")];
            tensor<string, []> x_321_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_321_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_17_begin_0 = const()[name = tensor<string, []>("k_cache_17_begin_0"), val = tensor<int32, [5]>([8, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_17_end_0 = const()[name = tensor<string, []>("k_cache_17_end_0"), val = tensor<int32, [5]>([9, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_17_end_mask_0 = const()[name = tensor<string, []>("k_cache_17_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_17_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_17_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_17_cast_fp16 = slice_by_index(begin = k_cache_17_begin_0, end = k_cache_17_end_0, end_mask = k_cache_17_end_mask_0, squeeze_mask = k_cache_17_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_17_cast_fp16")];
            tensor<int32, [5]> v_cache_17_begin_0 = const()[name = tensor<string, []>("v_cache_17_begin_0"), val = tensor<int32, [5]>([8, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_17_end_0 = const()[name = tensor<string, []>("v_cache_17_end_0"), val = tensor<int32, [5]>([9, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_17_end_mask_0 = const()[name = tensor<string, []>("v_cache_17_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_17_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_17_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_17_cast_fp16 = slice_by_index(begin = v_cache_17_begin_0, end = v_cache_17_end_0, end_mask = v_cache_17_end_mask_0, squeeze_mask = v_cache_17_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_17_cast_fp16")];
            tensor<int32, []> var_1523 = const()[name = tensor<string, []>("op_1523"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_1522_promoted = const()[name = tensor<string, []>("op_1522_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_321_cast_fp16_to_fp32 = cast(dtype = x_321_cast_fp16_to_fp32_dtype_0, x = x_321_cast_fp16)[name = tensor<string, []>("cast_542")];
            tensor<fp32, [1, 1, 512]> var_1532 = pow(x = x_321_cast_fp16_to_fp32, y = var_1522_promoted)[name = tensor<string, []>("op_1532")];
            tensor<int32, [1]> var_65_axes_0 = const()[name = tensor<string, []>("var_65_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_65_keep_dims_0 = const()[name = tensor<string, []>("var_65_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_65 = reduce_mean(axes = var_65_axes_0, keep_dims = var_65_keep_dims_0, x = var_1532)[name = tensor<string, []>("var_65")];
            tensor<string, []> var_65_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_65_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1536_to_fp16 = const()[name = tensor<string, []>("op_1536_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_65_to_fp16 = cast(dtype = var_65_to_fp16_dtype_0, x = var_65)[name = tensor<string, []>("cast_541")];
            tensor<fp16, [1, 1, 1]> var_1537_cast_fp16 = add(x = var_65_to_fp16, y = var_1536_to_fp16)[name = tensor<string, []>("op_1537_cast_fp16")];
            tensor<string, []> var_1537_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1537_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1538_epsilon_0 = const()[name = tensor<string, []>("op_1538_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1537_cast_fp16_to_fp32 = cast(dtype = var_1537_cast_fp16_to_fp32_dtype_0, x = var_1537_cast_fp16)[name = tensor<string, []>("cast_540")];
            tensor<fp32, [1, 1, 1]> var_1538 = rsqrt(epsilon = var_1538_epsilon_0, x = var_1537_cast_fp16_to_fp32)[name = tensor<string, []>("op_1538")];
            tensor<string, []> var_1538_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1538_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1538_to_fp16 = cast(dtype = var_1538_to_fp16_dtype_0, x = var_1538)[name = tensor<string, []>("cast_539")];
            tensor<fp16, [1, 1, 512]> x_327_cast_fp16 = mul(x = x_321_cast_fp16, y = var_1538_to_fp16)[name = tensor<string, []>("x_327_cast_fp16")];
            tensor<fp16, [512]> layers_8_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293787712)))];
            tensor<fp16, [1, 1, 512]> x_329_cast_fp16 = mul(x = layers_8_input_layernorm_weight_to_fp16, y = x_327_cast_fp16)[name = tensor<string, []>("x_329_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(293788800)))];
            tensor<fp16, [1, 1, 1536]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = x_329_cast_fp16)[name = tensor<string, []>("linear_56_cast_fp16")];
            tensor<int32, [4]> var_1552 = const()[name = tensor<string, []>("op_1552"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_331_cast_fp16 = reshape(shape = var_1552, x = linear_56_cast_fp16)[name = tensor<string, []>("x_331_cast_fp16")];
            tensor<string, []> x_331_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_331_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1522_promoted_1 = const()[name = tensor<string, []>("op_1522_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_331_cast_fp16_to_fp32 = cast(dtype = x_331_cast_fp16_to_fp32_dtype_0, x = x_331_cast_fp16)[name = tensor<string, []>("cast_538")];
            tensor<fp32, [1, 1, 12, 128]> var_1556 = pow(x = x_331_cast_fp16_to_fp32, y = var_1522_promoted_1)[name = tensor<string, []>("op_1556")];
            tensor<int32, [1]> var_67_axes_0 = const()[name = tensor<string, []>("var_67_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_67_keep_dims_0 = const()[name = tensor<string, []>("var_67_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_67 = reduce_mean(axes = var_67_axes_0, keep_dims = var_67_keep_dims_0, x = var_1556)[name = tensor<string, []>("var_67")];
            tensor<string, []> var_67_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_67_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1560_to_fp16 = const()[name = tensor<string, []>("op_1560_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_67_to_fp16 = cast(dtype = var_67_to_fp16_dtype_0, x = var_67)[name = tensor<string, []>("cast_537")];
            tensor<fp16, [1, 1, 12, 1]> var_1561_cast_fp16 = add(x = var_67_to_fp16, y = var_1560_to_fp16)[name = tensor<string, []>("op_1561_cast_fp16")];
            tensor<string, []> var_1561_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1561_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1562_epsilon_0 = const()[name = tensor<string, []>("op_1562_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_1561_cast_fp16_to_fp32 = cast(dtype = var_1561_cast_fp16_to_fp32_dtype_0, x = var_1561_cast_fp16)[name = tensor<string, []>("cast_536")];
            tensor<fp32, [1, 1, 12, 1]> var_1562 = rsqrt(epsilon = var_1562_epsilon_0, x = var_1561_cast_fp16_to_fp32)[name = tensor<string, []>("op_1562")];
            tensor<string, []> var_1562_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1562_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_1562_to_fp16 = cast(dtype = var_1562_to_fp16_dtype_0, x = var_1562)[name = tensor<string, []>("cast_535")];
            tensor<fp16, [1, 1, 12, 128]> x_337_cast_fp16 = mul(x = x_331_cast_fp16, y = var_1562_to_fp16)[name = tensor<string, []>("x_337_cast_fp16")];
            tensor<fp16, [128]> layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295361728)))];
            tensor<fp16, [1, 1, 12, 128]> var_1564_cast_fp16 = mul(x = layers_8_self_attn_q_norm_weight_to_fp16, y = x_337_cast_fp16)[name = tensor<string, []>("op_1564_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = tensor<string, []>("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295362048)))];
            tensor<fp16, [1, 1, 512]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = x_329_cast_fp16)[name = tensor<string, []>("linear_57_cast_fp16")];
            tensor<int32, [4]> var_1568 = const()[name = tensor<string, []>("op_1568"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_339_cast_fp16 = reshape(shape = var_1568, x = linear_57_cast_fp16)[name = tensor<string, []>("x_339_cast_fp16")];
            tensor<string, []> x_339_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_339_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1522_promoted_2 = const()[name = tensor<string, []>("op_1522_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_339_cast_fp16_to_fp32 = cast(dtype = x_339_cast_fp16_to_fp32_dtype_0, x = x_339_cast_fp16)[name = tensor<string, []>("cast_534")];
            tensor<fp32, [1, 1, 4, 128]> var_1572 = pow(x = x_339_cast_fp16_to_fp32, y = var_1522_promoted_2)[name = tensor<string, []>("op_1572")];
            tensor<int32, [1]> var_69_axes_0 = const()[name = tensor<string, []>("var_69_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_69_keep_dims_0 = const()[name = tensor<string, []>("var_69_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_69 = reduce_mean(axes = var_69_axes_0, keep_dims = var_69_keep_dims_0, x = var_1572)[name = tensor<string, []>("var_69")];
            tensor<string, []> var_69_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_69_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1576_to_fp16 = const()[name = tensor<string, []>("op_1576_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_69_to_fp16 = cast(dtype = var_69_to_fp16_dtype_0, x = var_69)[name = tensor<string, []>("cast_533")];
            tensor<fp16, [1, 1, 4, 1]> var_1577_cast_fp16 = add(x = var_69_to_fp16, y = var_1576_to_fp16)[name = tensor<string, []>("op_1577_cast_fp16")];
            tensor<string, []> var_1577_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1577_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1578_epsilon_0 = const()[name = tensor<string, []>("op_1578_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_1577_cast_fp16_to_fp32 = cast(dtype = var_1577_cast_fp16_to_fp32_dtype_0, x = var_1577_cast_fp16)[name = tensor<string, []>("cast_532")];
            tensor<fp32, [1, 1, 4, 1]> var_1578 = rsqrt(epsilon = var_1578_epsilon_0, x = var_1577_cast_fp16_to_fp32)[name = tensor<string, []>("op_1578")];
            tensor<string, []> var_1578_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1578_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_1578_to_fp16 = cast(dtype = var_1578_to_fp16_dtype_0, x = var_1578)[name = tensor<string, []>("cast_531")];
            tensor<fp16, [1, 1, 4, 128]> x_345_cast_fp16 = mul(x = x_339_cast_fp16, y = var_1578_to_fp16)[name = tensor<string, []>("x_345_cast_fp16")];
            tensor<fp16, [128]> layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295886400)))];
            tensor<fp16, [1, 1, 4, 128]> var_1580_cast_fp16 = mul(x = layers_8_self_attn_k_norm_weight_to_fp16, y = x_345_cast_fp16)[name = tensor<string, []>("op_1580_cast_fp16")];
            tensor<int32, [4]> k_33_perm_0 = const()[name = tensor<string, []>("k_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295886720)))];
            tensor<fp16, [1, 1, 512]> linear_58_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = x_329_cast_fp16)[name = tensor<string, []>("linear_58_cast_fp16")];
            tensor<int32, [4]> var_1584 = const()[name = tensor<string, []>("op_1584"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_1585_cast_fp16 = reshape(shape = var_1584, x = linear_58_cast_fp16)[name = tensor<string, []>("op_1585_cast_fp16")];
            tensor<int32, [4]> v_17_perm_0 = const()[name = tensor<string, []>("v_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = var_1564_cast_fp16)[name = tensor<string, []>("transpose_79")];
            tensor<fp16, [1, 12, 1, 128]> var_1589_cast_fp16 = mul(x = q_33_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1589_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = tensor<string, []>("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = tensor<string, []>("x1_33_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = tensor<string, []>("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_33_cast_fp16)[name = tensor<string, []>("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = tensor<string, []>("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = tensor<string, []>("x2_33_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = tensor<string, []>("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_33_cast_fp16)[name = tensor<string, []>("x2_33_cast_fp16")];
            tensor<fp16, []> const_17_promoted_to_fp16 = const()[name = tensor<string, []>("const_17_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_1592_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_17_promoted_to_fp16)[name = tensor<string, []>("op_1592_cast_fp16")];
            tensor<bool, []> var_1594_interleave_0 = const()[name = tensor<string, []>("op_1594_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_1594_cast_fp16 = concat(axis = var_1523, interleave = var_1594_interleave_0, values = (var_1592_cast_fp16, x1_33_cast_fp16))[name = tensor<string, []>("op_1594_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_1595_cast_fp16 = mul(x = var_1594_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1595_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_35_cast_fp16 = add(x = var_1589_cast_fp16, y = var_1595_cast_fp16)[name = tensor<string, []>("q_35_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_33_cast_fp16 = transpose(perm = k_33_perm_0, x = var_1580_cast_fp16)[name = tensor<string, []>("transpose_78")];
            tensor<fp16, [1, 4, 1, 128]> var_1597_cast_fp16 = mul(x = k_33_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1597_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = tensor<string, []>("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = tensor<string, []>("x1_35_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = tensor<string, []>("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_33_cast_fp16)[name = tensor<string, []>("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = tensor<string, []>("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = tensor<string, []>("x2_35_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = tensor<string, []>("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_33_cast_fp16)[name = tensor<string, []>("x2_35_cast_fp16")];
            tensor<fp16, []> const_18_promoted_to_fp16 = const()[name = tensor<string, []>("const_18_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_1600_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_18_promoted_to_fp16)[name = tensor<string, []>("op_1600_cast_fp16")];
            tensor<bool, []> var_1602_interleave_0 = const()[name = tensor<string, []>("op_1602_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_1602_cast_fp16 = concat(axis = var_1523, interleave = var_1602_interleave_0, values = (var_1600_cast_fp16, x1_35_cast_fp16))[name = tensor<string, []>("op_1602_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_1603_cast_fp16 = mul(x = var_1602_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1603_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_35_cast_fp16 = add(x = var_1597_cast_fp16, y = var_1603_cast_fp16)[name = tensor<string, []>("k_35_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1606_cast_fp16 = mul(x = k_cache_17_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1606_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1607_cast_fp16 = mul(x = k_35_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1607_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_17_cast_fp16 = add(x = var_1606_cast_fp16, y = var_1607_cast_fp16)[name = tensor<string, []>("k_full_17_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1610_cast_fp16 = mul(x = v_cache_17_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1610_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = var_1585_cast_fp16)[name = tensor<string, []>("transpose_77")];
            tensor<fp16, [1, 4, 2048, 128]> var_1611_cast_fp16 = mul(x = v_17_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1611_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_17_cast_fp16 = add(x = var_1610_cast_fp16, y = var_1611_cast_fp16)[name = tensor<string, []>("v_full_17_cast_fp16")];
            tensor<int32, [1]> var_1613_axes_0 = const()[name = tensor<string, []>("op_1613_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1613_cast_fp16 = expand_dims(axes = var_1613_axes_0, x = k_full_17_cast_fp16)[name = tensor<string, []>("op_1613_cast_fp16")];
            tensor<int32, [5]> var_1615_reps_0 = const()[name = tensor<string, []>("op_1615_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1615_cast_fp16 = tile(reps = var_1615_reps_0, x = var_1613_cast_fp16)[name = tensor<string, []>("op_1615_cast_fp16")];
            tensor<int32, [4]> var_1616 = const()[name = tensor<string, []>("op_1616"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_17_cast_fp16 = reshape(shape = var_1616, x = var_1615_cast_fp16)[name = tensor<string, []>("k_rep_17_cast_fp16")];
            tensor<int32, [1]> var_1618_axes_0 = const()[name = tensor<string, []>("op_1618_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1618_cast_fp16 = expand_dims(axes = var_1618_axes_0, x = v_full_17_cast_fp16)[name = tensor<string, []>("op_1618_cast_fp16")];
            tensor<int32, [5]> var_1620_reps_0 = const()[name = tensor<string, []>("op_1620_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1620_cast_fp16 = tile(reps = var_1620_reps_0, x = var_1618_cast_fp16)[name = tensor<string, []>("op_1620_cast_fp16")];
            tensor<int32, [4]> var_1621 = const()[name = tensor<string, []>("op_1621"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_17_cast_fp16 = reshape(shape = var_1621, x = var_1620_cast_fp16)[name = tensor<string, []>("v_rep_17_cast_fp16")];
            tensor<bool, []> var_1624_transpose_x_1 = const()[name = tensor<string, []>("op_1624_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1624_transpose_y_1 = const()[name = tensor<string, []>("op_1624_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_1624_cast_fp16 = matmul(transpose_x = var_1624_transpose_x_1, transpose_y = var_1624_transpose_y_1, x = q_35_cast_fp16, y = k_rep_17_cast_fp16)[name = tensor<string, []>("op_1624_cast_fp16")];
            tensor<fp16, []> var_1625_to_fp16 = const()[name = tensor<string, []>("op_1625_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_33_cast_fp16 = mul(x = var_1624_cast_fp16, y = var_1625_to_fp16)[name = tensor<string, []>("attn_33_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_35_cast_fp16 = add(x = attn_33_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_35_cast_fp16")];
            tensor<string, []> input_35_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_35_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_35_cast_fp16_to_fp32 = cast(dtype = input_35_cast_fp16_to_fp32_dtype_0, x = input_35_cast_fp16)[name = tensor<string, []>("cast_530")];
            tensor<fp32, [1, 12, 1, 2048]> attn_35 = softmax(axis = var_1523, x = input_35_cast_fp16_to_fp32)[name = tensor<string, []>("attn_35")];
            tensor<bool, []> out_17_transpose_x_0 = const()[name = tensor<string, []>("out_17_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_17_transpose_y_0 = const()[name = tensor<string, []>("out_17_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_35_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_35_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_35_to_fp16 = cast(dtype = attn_35_to_fp16_dtype_0, x = attn_35)[name = tensor<string, []>("cast_529")];
            tensor<fp16, [1, 12, 1, 128]> out_17_cast_fp16 = matmul(transpose_x = out_17_transpose_x_0, transpose_y = out_17_transpose_y_0, x = attn_35_to_fp16, y = v_rep_17_cast_fp16)[name = tensor<string, []>("out_17_cast_fp16")];
            tensor<int32, [4]> var_1630_perm_0 = const()[name = tensor<string, []>("op_1630_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1632 = const()[name = tensor<string, []>("op_1632"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_1630_cast_fp16 = transpose(perm = var_1630_perm_0, x = out_17_cast_fp16)[name = tensor<string, []>("transpose_76")];
            tensor<fp16, [1, 1, 1536]> x_347_cast_fp16 = reshape(shape = var_1632, x = var_1630_cast_fp16)[name = tensor<string, []>("x_347_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(296411072)))];
            tensor<fp16, [1, 1, 512]> linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = x_347_cast_fp16)[name = tensor<string, []>("linear_59_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_349_cast_fp16 = add(x = x_321_cast_fp16, y = linear_59_cast_fp16)[name = tensor<string, []>("x_349_cast_fp16")];
            tensor<string, []> x_349_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_349_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1522_promoted_3 = const()[name = tensor<string, []>("op_1522_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_349_cast_fp16_to_fp32 = cast(dtype = x_349_cast_fp16_to_fp32_dtype_0, x = x_349_cast_fp16)[name = tensor<string, []>("cast_528")];
            tensor<fp32, [1, 1, 512]> var_1643 = pow(x = x_349_cast_fp16_to_fp32, y = var_1522_promoted_3)[name = tensor<string, []>("op_1643")];
            tensor<int32, [1]> var_71_axes_0 = const()[name = tensor<string, []>("var_71_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_71_keep_dims_0 = const()[name = tensor<string, []>("var_71_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_71 = reduce_mean(axes = var_71_axes_0, keep_dims = var_71_keep_dims_0, x = var_1643)[name = tensor<string, []>("var_71")];
            tensor<string, []> var_71_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_71_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1647_to_fp16 = const()[name = tensor<string, []>("op_1647_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_71_to_fp16 = cast(dtype = var_71_to_fp16_dtype_0, x = var_71)[name = tensor<string, []>("cast_527")];
            tensor<fp16, [1, 1, 1]> var_1648_cast_fp16 = add(x = var_71_to_fp16, y = var_1647_to_fp16)[name = tensor<string, []>("op_1648_cast_fp16")];
            tensor<string, []> var_1648_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1648_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1649_epsilon_0 = const()[name = tensor<string, []>("op_1649_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1648_cast_fp16_to_fp32 = cast(dtype = var_1648_cast_fp16_to_fp32_dtype_0, x = var_1648_cast_fp16)[name = tensor<string, []>("cast_526")];
            tensor<fp32, [1, 1, 1]> var_1649 = rsqrt(epsilon = var_1649_epsilon_0, x = var_1648_cast_fp16_to_fp32)[name = tensor<string, []>("op_1649")];
            tensor<string, []> var_1649_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1649_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1649_to_fp16 = cast(dtype = var_1649_to_fp16_dtype_0, x = var_1649)[name = tensor<string, []>("cast_525")];
            tensor<fp16, [1, 1, 512]> x_355_cast_fp16 = mul(x = x_349_cast_fp16, y = var_1649_to_fp16)[name = tensor<string, []>("x_355_cast_fp16")];
            tensor<fp16, [512]> layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297984000)))];
            tensor<fp16, [1, 1, 512]> x_357_cast_fp16 = mul(x = layers_8_post_attention_layernorm_weight_to_fp16, y = x_355_cast_fp16)[name = tensor<string, []>("x_357_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_8_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(297985088)))];
            tensor<fp16, [1, 1, 1536]> linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_mlp_gate_proj_weight_to_fp16, x = x_357_cast_fp16)[name = tensor<string, []>("linear_60_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_1660_cast_fp16 = silu(x = linear_60_cast_fp16)[name = tensor<string, []>("op_1660_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_8_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(299558016)))];
            tensor<fp16, [1, 1, 1536]> linear_61_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_mlp_up_proj_weight_to_fp16, x = x_357_cast_fp16)[name = tensor<string, []>("linear_61_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_359_cast_fp16 = mul(x = var_1660_cast_fp16, y = linear_61_cast_fp16)[name = tensor<string, []>("x_359_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_8_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_8_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(301130944)))];
            tensor<fp16, [1, 1, 512]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_8_mlp_down_proj_weight_to_fp16, x = x_359_cast_fp16)[name = tensor<string, []>("linear_62_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_361_cast_fp16 = add(x = x_349_cast_fp16, y = linear_62_cast_fp16)[name = tensor<string, []>("x_361_cast_fp16")];
            tensor<string, []> x_361_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_361_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_19_begin_0 = const()[name = tensor<string, []>("k_cache_19_begin_0"), val = tensor<int32, [5]>([9, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_19_end_0 = const()[name = tensor<string, []>("k_cache_19_end_0"), val = tensor<int32, [5]>([10, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_19_end_mask_0 = const()[name = tensor<string, []>("k_cache_19_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_19_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_19_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_19_cast_fp16 = slice_by_index(begin = k_cache_19_begin_0, end = k_cache_19_end_0, end_mask = k_cache_19_end_mask_0, squeeze_mask = k_cache_19_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_19_cast_fp16")];
            tensor<int32, [5]> v_cache_19_begin_0 = const()[name = tensor<string, []>("v_cache_19_begin_0"), val = tensor<int32, [5]>([9, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_19_end_0 = const()[name = tensor<string, []>("v_cache_19_end_0"), val = tensor<int32, [5]>([10, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_19_end_mask_0 = const()[name = tensor<string, []>("v_cache_19_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_19_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_19_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_19_cast_fp16 = slice_by_index(begin = v_cache_19_begin_0, end = v_cache_19_end_0, end_mask = v_cache_19_end_mask_0, squeeze_mask = v_cache_19_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_19_cast_fp16")];
            tensor<int32, []> var_1692 = const()[name = tensor<string, []>("op_1692"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_1691_promoted = const()[name = tensor<string, []>("op_1691_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_361_cast_fp16_to_fp32 = cast(dtype = x_361_cast_fp16_to_fp32_dtype_0, x = x_361_cast_fp16)[name = tensor<string, []>("cast_524")];
            tensor<fp32, [1, 1, 512]> var_1701 = pow(x = x_361_cast_fp16_to_fp32, y = var_1691_promoted)[name = tensor<string, []>("op_1701")];
            tensor<int32, [1]> var_73_axes_0 = const()[name = tensor<string, []>("var_73_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_73_keep_dims_0 = const()[name = tensor<string, []>("var_73_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_73 = reduce_mean(axes = var_73_axes_0, keep_dims = var_73_keep_dims_0, x = var_1701)[name = tensor<string, []>("var_73")];
            tensor<string, []> var_73_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_73_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1705_to_fp16 = const()[name = tensor<string, []>("op_1705_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_73_to_fp16 = cast(dtype = var_73_to_fp16_dtype_0, x = var_73)[name = tensor<string, []>("cast_523")];
            tensor<fp16, [1, 1, 1]> var_1706_cast_fp16 = add(x = var_73_to_fp16, y = var_1705_to_fp16)[name = tensor<string, []>("op_1706_cast_fp16")];
            tensor<string, []> var_1706_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1706_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1707_epsilon_0 = const()[name = tensor<string, []>("op_1707_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1706_cast_fp16_to_fp32 = cast(dtype = var_1706_cast_fp16_to_fp32_dtype_0, x = var_1706_cast_fp16)[name = tensor<string, []>("cast_522")];
            tensor<fp32, [1, 1, 1]> var_1707 = rsqrt(epsilon = var_1707_epsilon_0, x = var_1706_cast_fp16_to_fp32)[name = tensor<string, []>("op_1707")];
            tensor<string, []> var_1707_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1707_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1707_to_fp16 = cast(dtype = var_1707_to_fp16_dtype_0, x = var_1707)[name = tensor<string, []>("cast_521")];
            tensor<fp16, [1, 1, 512]> x_367_cast_fp16 = mul(x = x_361_cast_fp16, y = var_1707_to_fp16)[name = tensor<string, []>("x_367_cast_fp16")];
            tensor<fp16, [512]> layers_9_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302703872)))];
            tensor<fp16, [1, 1, 512]> x_369_cast_fp16 = mul(x = layers_9_input_layernorm_weight_to_fp16, y = x_367_cast_fp16)[name = tensor<string, []>("x_369_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302704960)))];
            tensor<fp16, [1, 1, 1536]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = x_369_cast_fp16)[name = tensor<string, []>("linear_63_cast_fp16")];
            tensor<int32, [4]> var_1721 = const()[name = tensor<string, []>("op_1721"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_371_cast_fp16 = reshape(shape = var_1721, x = linear_63_cast_fp16)[name = tensor<string, []>("x_371_cast_fp16")];
            tensor<string, []> x_371_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_371_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1691_promoted_1 = const()[name = tensor<string, []>("op_1691_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_371_cast_fp16_to_fp32 = cast(dtype = x_371_cast_fp16_to_fp32_dtype_0, x = x_371_cast_fp16)[name = tensor<string, []>("cast_520")];
            tensor<fp32, [1, 1, 12, 128]> var_1725 = pow(x = x_371_cast_fp16_to_fp32, y = var_1691_promoted_1)[name = tensor<string, []>("op_1725")];
            tensor<int32, [1]> var_75_axes_0 = const()[name = tensor<string, []>("var_75_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_75_keep_dims_0 = const()[name = tensor<string, []>("var_75_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_75_0 = reduce_mean(axes = var_75_axes_0, keep_dims = var_75_keep_dims_0, x = var_1725)[name = tensor<string, []>("var_75")];
            tensor<string, []> var_75_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_75_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1729_to_fp16 = const()[name = tensor<string, []>("op_1729_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_75_to_fp16 = cast(dtype = var_75_to_fp16_dtype_0, x = var_75_0)[name = tensor<string, []>("cast_519")];
            tensor<fp16, [1, 1, 12, 1]> var_1730_cast_fp16 = add(x = var_75_to_fp16, y = var_1729_to_fp16)[name = tensor<string, []>("op_1730_cast_fp16")];
            tensor<string, []> var_1730_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1730_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1731_epsilon_0 = const()[name = tensor<string, []>("op_1731_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_1730_cast_fp16_to_fp32 = cast(dtype = var_1730_cast_fp16_to_fp32_dtype_0, x = var_1730_cast_fp16)[name = tensor<string, []>("cast_518")];
            tensor<fp32, [1, 1, 12, 1]> var_1731 = rsqrt(epsilon = var_1731_epsilon_0, x = var_1730_cast_fp16_to_fp32)[name = tensor<string, []>("op_1731")];
            tensor<string, []> var_1731_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1731_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_1731_to_fp16 = cast(dtype = var_1731_to_fp16_dtype_0, x = var_1731)[name = tensor<string, []>("cast_517")];
            tensor<fp16, [1, 1, 12, 128]> x_377_cast_fp16 = mul(x = x_371_cast_fp16, y = var_1731_to_fp16)[name = tensor<string, []>("x_377_cast_fp16")];
            tensor<fp16, [128]> layers_9_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304277888)))];
            tensor<fp16, [1, 1, 12, 128]> var_1733_cast_fp16 = mul(x = layers_9_self_attn_q_norm_weight_to_fp16, y = x_377_cast_fp16)[name = tensor<string, []>("op_1733_cast_fp16")];
            tensor<int32, [4]> q_37_perm_0 = const()[name = tensor<string, []>("q_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304278208)))];
            tensor<fp16, [1, 1, 512]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = x_369_cast_fp16)[name = tensor<string, []>("linear_64_cast_fp16")];
            tensor<int32, [4]> var_1737 = const()[name = tensor<string, []>("op_1737"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_379_cast_fp16 = reshape(shape = var_1737, x = linear_64_cast_fp16)[name = tensor<string, []>("x_379_cast_fp16")];
            tensor<string, []> x_379_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_379_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1691_promoted_2 = const()[name = tensor<string, []>("op_1691_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_379_cast_fp16_to_fp32 = cast(dtype = x_379_cast_fp16_to_fp32_dtype_0, x = x_379_cast_fp16)[name = tensor<string, []>("cast_516")];
            tensor<fp32, [1, 1, 4, 128]> var_1741 = pow(x = x_379_cast_fp16_to_fp32, y = var_1691_promoted_2)[name = tensor<string, []>("op_1741")];
            tensor<int32, [1]> var_77_axes_0 = const()[name = tensor<string, []>("var_77_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_77_keep_dims_0 = const()[name = tensor<string, []>("var_77_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_77 = reduce_mean(axes = var_77_axes_0, keep_dims = var_77_keep_dims_0, x = var_1741)[name = tensor<string, []>("var_77")];
            tensor<string, []> var_77_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_77_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1745_to_fp16 = const()[name = tensor<string, []>("op_1745_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_77_to_fp16 = cast(dtype = var_77_to_fp16_dtype_0, x = var_77)[name = tensor<string, []>("cast_515")];
            tensor<fp16, [1, 1, 4, 1]> var_1746_cast_fp16 = add(x = var_77_to_fp16, y = var_1745_to_fp16)[name = tensor<string, []>("op_1746_cast_fp16")];
            tensor<string, []> var_1746_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1746_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1747_epsilon_0 = const()[name = tensor<string, []>("op_1747_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_1746_cast_fp16_to_fp32 = cast(dtype = var_1746_cast_fp16_to_fp32_dtype_0, x = var_1746_cast_fp16)[name = tensor<string, []>("cast_514")];
            tensor<fp32, [1, 1, 4, 1]> var_1747 = rsqrt(epsilon = var_1747_epsilon_0, x = var_1746_cast_fp16_to_fp32)[name = tensor<string, []>("op_1747")];
            tensor<string, []> var_1747_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1747_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_1747_to_fp16 = cast(dtype = var_1747_to_fp16_dtype_0, x = var_1747)[name = tensor<string, []>("cast_513")];
            tensor<fp16, [1, 1, 4, 128]> x_385_cast_fp16 = mul(x = x_379_cast_fp16, y = var_1747_to_fp16)[name = tensor<string, []>("x_385_cast_fp16")];
            tensor<fp16, [128]> layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304802560)))];
            tensor<fp16, [1, 1, 4, 128]> var_1749_cast_fp16 = mul(x = layers_9_self_attn_k_norm_weight_to_fp16, y = x_385_cast_fp16)[name = tensor<string, []>("op_1749_cast_fp16")];
            tensor<int32, [4]> k_37_perm_0 = const()[name = tensor<string, []>("k_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304802880)))];
            tensor<fp16, [1, 1, 512]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = x_369_cast_fp16)[name = tensor<string, []>("linear_65_cast_fp16")];
            tensor<int32, [4]> var_1753 = const()[name = tensor<string, []>("op_1753"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_1754_cast_fp16 = reshape(shape = var_1753, x = linear_65_cast_fp16)[name = tensor<string, []>("op_1754_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = tensor<string, []>("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_37_cast_fp16 = transpose(perm = q_37_perm_0, x = var_1733_cast_fp16)[name = tensor<string, []>("transpose_75")];
            tensor<fp16, [1, 12, 1, 128]> var_1758_cast_fp16 = mul(x = q_37_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1758_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = tensor<string, []>("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = tensor<string, []>("x1_37_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = tensor<string, []>("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_37_cast_fp16)[name = tensor<string, []>("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = tensor<string, []>("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = tensor<string, []>("x2_37_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = tensor<string, []>("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_37_cast_fp16)[name = tensor<string, []>("x2_37_cast_fp16")];
            tensor<fp16, []> const_19_promoted_to_fp16 = const()[name = tensor<string, []>("const_19_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_1761_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_19_promoted_to_fp16)[name = tensor<string, []>("op_1761_cast_fp16")];
            tensor<bool, []> var_1763_interleave_0 = const()[name = tensor<string, []>("op_1763_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_1763_cast_fp16 = concat(axis = var_1692, interleave = var_1763_interleave_0, values = (var_1761_cast_fp16, x1_37_cast_fp16))[name = tensor<string, []>("op_1763_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_1764_cast_fp16 = mul(x = var_1763_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1764_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_39_cast_fp16 = add(x = var_1758_cast_fp16, y = var_1764_cast_fp16)[name = tensor<string, []>("q_39_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_37_cast_fp16 = transpose(perm = k_37_perm_0, x = var_1749_cast_fp16)[name = tensor<string, []>("transpose_74")];
            tensor<fp16, [1, 4, 1, 128]> var_1766_cast_fp16 = mul(x = k_37_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1766_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = tensor<string, []>("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = tensor<string, []>("x1_39_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = tensor<string, []>("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_37_cast_fp16)[name = tensor<string, []>("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = tensor<string, []>("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = tensor<string, []>("x2_39_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = tensor<string, []>("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_37_cast_fp16)[name = tensor<string, []>("x2_39_cast_fp16")];
            tensor<fp16, []> const_20_promoted_to_fp16 = const()[name = tensor<string, []>("const_20_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_1769_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_20_promoted_to_fp16)[name = tensor<string, []>("op_1769_cast_fp16")];
            tensor<bool, []> var_1771_interleave_0 = const()[name = tensor<string, []>("op_1771_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_1771_cast_fp16 = concat(axis = var_1692, interleave = var_1771_interleave_0, values = (var_1769_cast_fp16, x1_39_cast_fp16))[name = tensor<string, []>("op_1771_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_1772_cast_fp16 = mul(x = var_1771_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1772_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_39_cast_fp16 = add(x = var_1766_cast_fp16, y = var_1772_cast_fp16)[name = tensor<string, []>("k_39_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1775_cast_fp16 = mul(x = k_cache_19_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1775_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1776_cast_fp16 = mul(x = k_39_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1776_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_19_cast_fp16 = add(x = var_1775_cast_fp16, y = var_1776_cast_fp16)[name = tensor<string, []>("k_full_19_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1779_cast_fp16 = mul(x = v_cache_19_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1779_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_1754_cast_fp16)[name = tensor<string, []>("transpose_73")];
            tensor<fp16, [1, 4, 2048, 128]> var_1780_cast_fp16 = mul(x = v_19_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1780_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_19_cast_fp16 = add(x = var_1779_cast_fp16, y = var_1780_cast_fp16)[name = tensor<string, []>("v_full_19_cast_fp16")];
            tensor<int32, [1]> var_1782_axes_0 = const()[name = tensor<string, []>("op_1782_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1782_cast_fp16 = expand_dims(axes = var_1782_axes_0, x = k_full_19_cast_fp16)[name = tensor<string, []>("op_1782_cast_fp16")];
            tensor<int32, [5]> var_1784_reps_0 = const()[name = tensor<string, []>("op_1784_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1784_cast_fp16 = tile(reps = var_1784_reps_0, x = var_1782_cast_fp16)[name = tensor<string, []>("op_1784_cast_fp16")];
            tensor<int32, [4]> var_1785 = const()[name = tensor<string, []>("op_1785"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_19_cast_fp16 = reshape(shape = var_1785, x = var_1784_cast_fp16)[name = tensor<string, []>("k_rep_19_cast_fp16")];
            tensor<int32, [1]> var_1787_axes_0 = const()[name = tensor<string, []>("op_1787_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1787_cast_fp16 = expand_dims(axes = var_1787_axes_0, x = v_full_19_cast_fp16)[name = tensor<string, []>("op_1787_cast_fp16")];
            tensor<int32, [5]> var_1789_reps_0 = const()[name = tensor<string, []>("op_1789_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1789_cast_fp16 = tile(reps = var_1789_reps_0, x = var_1787_cast_fp16)[name = tensor<string, []>("op_1789_cast_fp16")];
            tensor<int32, [4]> var_1790 = const()[name = tensor<string, []>("op_1790"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_19_cast_fp16 = reshape(shape = var_1790, x = var_1789_cast_fp16)[name = tensor<string, []>("v_rep_19_cast_fp16")];
            tensor<bool, []> var_1793_transpose_x_1 = const()[name = tensor<string, []>("op_1793_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1793_transpose_y_1 = const()[name = tensor<string, []>("op_1793_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_1793_cast_fp16 = matmul(transpose_x = var_1793_transpose_x_1, transpose_y = var_1793_transpose_y_1, x = q_39_cast_fp16, y = k_rep_19_cast_fp16)[name = tensor<string, []>("op_1793_cast_fp16")];
            tensor<fp16, []> var_1794_to_fp16 = const()[name = tensor<string, []>("op_1794_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_37_cast_fp16 = mul(x = var_1793_cast_fp16, y = var_1794_to_fp16)[name = tensor<string, []>("attn_37_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_39_cast_fp16 = add(x = attn_37_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_39_cast_fp16")];
            tensor<string, []> input_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_39_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_39_cast_fp16_to_fp32 = cast(dtype = input_39_cast_fp16_to_fp32_dtype_0, x = input_39_cast_fp16)[name = tensor<string, []>("cast_512")];
            tensor<fp32, [1, 12, 1, 2048]> attn_39 = softmax(axis = var_1692, x = input_39_cast_fp16_to_fp32)[name = tensor<string, []>("attn_39")];
            tensor<bool, []> out_19_transpose_x_0 = const()[name = tensor<string, []>("out_19_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_19_transpose_y_0 = const()[name = tensor<string, []>("out_19_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_39_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_39_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_39_to_fp16 = cast(dtype = attn_39_to_fp16_dtype_0, x = attn_39)[name = tensor<string, []>("cast_511")];
            tensor<fp16, [1, 12, 1, 128]> out_19_cast_fp16 = matmul(transpose_x = out_19_transpose_x_0, transpose_y = out_19_transpose_y_0, x = attn_39_to_fp16, y = v_rep_19_cast_fp16)[name = tensor<string, []>("out_19_cast_fp16")];
            tensor<int32, [4]> var_1799_perm_0 = const()[name = tensor<string, []>("op_1799_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1801 = const()[name = tensor<string, []>("op_1801"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_1799_cast_fp16 = transpose(perm = var_1799_perm_0, x = out_19_cast_fp16)[name = tensor<string, []>("transpose_72")];
            tensor<fp16, [1, 1, 1536]> x_387_cast_fp16 = reshape(shape = var_1801, x = var_1799_cast_fp16)[name = tensor<string, []>("x_387_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305327232)))];
            tensor<fp16, [1, 1, 512]> linear_66_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = x_387_cast_fp16)[name = tensor<string, []>("linear_66_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_389_cast_fp16 = add(x = x_361_cast_fp16, y = linear_66_cast_fp16)[name = tensor<string, []>("x_389_cast_fp16")];
            tensor<string, []> x_389_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_389_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1691_promoted_3 = const()[name = tensor<string, []>("op_1691_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_389_cast_fp16_to_fp32 = cast(dtype = x_389_cast_fp16_to_fp32_dtype_0, x = x_389_cast_fp16)[name = tensor<string, []>("cast_510")];
            tensor<fp32, [1, 1, 512]> var_1812 = pow(x = x_389_cast_fp16_to_fp32, y = var_1691_promoted_3)[name = tensor<string, []>("op_1812")];
            tensor<int32, [1]> var_79_axes_0 = const()[name = tensor<string, []>("var_79_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_79_keep_dims_0 = const()[name = tensor<string, []>("var_79_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_79 = reduce_mean(axes = var_79_axes_0, keep_dims = var_79_keep_dims_0, x = var_1812)[name = tensor<string, []>("var_79")];
            tensor<string, []> var_79_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_79_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1816_to_fp16 = const()[name = tensor<string, []>("op_1816_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_79_to_fp16 = cast(dtype = var_79_to_fp16_dtype_0, x = var_79)[name = tensor<string, []>("cast_509")];
            tensor<fp16, [1, 1, 1]> var_1817_cast_fp16 = add(x = var_79_to_fp16, y = var_1816_to_fp16)[name = tensor<string, []>("op_1817_cast_fp16")];
            tensor<string, []> var_1817_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1817_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1818_epsilon_0 = const()[name = tensor<string, []>("op_1818_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1817_cast_fp16_to_fp32 = cast(dtype = var_1817_cast_fp16_to_fp32_dtype_0, x = var_1817_cast_fp16)[name = tensor<string, []>("cast_508")];
            tensor<fp32, [1, 1, 1]> var_1818 = rsqrt(epsilon = var_1818_epsilon_0, x = var_1817_cast_fp16_to_fp32)[name = tensor<string, []>("op_1818")];
            tensor<string, []> var_1818_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1818_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1818_to_fp16 = cast(dtype = var_1818_to_fp16_dtype_0, x = var_1818)[name = tensor<string, []>("cast_507")];
            tensor<fp16, [1, 1, 512]> x_395_cast_fp16 = mul(x = x_389_cast_fp16, y = var_1818_to_fp16)[name = tensor<string, []>("x_395_cast_fp16")];
            tensor<fp16, [512]> layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306900160)))];
            tensor<fp16, [1, 1, 512]> x_397_cast_fp16 = mul(x = layers_9_post_attention_layernorm_weight_to_fp16, y = x_395_cast_fp16)[name = tensor<string, []>("x_397_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_9_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306901248)))];
            tensor<fp16, [1, 1, 1536]> linear_67_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_mlp_gate_proj_weight_to_fp16, x = x_397_cast_fp16)[name = tensor<string, []>("linear_67_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_1829_cast_fp16 = silu(x = linear_67_cast_fp16)[name = tensor<string, []>("op_1829_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_9_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308474176)))];
            tensor<fp16, [1, 1, 1536]> linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_mlp_up_proj_weight_to_fp16, x = x_397_cast_fp16)[name = tensor<string, []>("linear_68_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_399_cast_fp16 = mul(x = var_1829_cast_fp16, y = linear_68_cast_fp16)[name = tensor<string, []>("x_399_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_9_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_9_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310047104)))];
            tensor<fp16, [1, 1, 512]> linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_9_mlp_down_proj_weight_to_fp16, x = x_399_cast_fp16)[name = tensor<string, []>("linear_69_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_401_cast_fp16 = add(x = x_389_cast_fp16, y = linear_69_cast_fp16)[name = tensor<string, []>("x_401_cast_fp16")];
            tensor<string, []> x_401_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_401_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_21_begin_0 = const()[name = tensor<string, []>("k_cache_21_begin_0"), val = tensor<int32, [5]>([10, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_21_end_0 = const()[name = tensor<string, []>("k_cache_21_end_0"), val = tensor<int32, [5]>([11, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_21_end_mask_0 = const()[name = tensor<string, []>("k_cache_21_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_21_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_21_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_21_cast_fp16 = slice_by_index(begin = k_cache_21_begin_0, end = k_cache_21_end_0, end_mask = k_cache_21_end_mask_0, squeeze_mask = k_cache_21_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_21_cast_fp16")];
            tensor<int32, [5]> v_cache_21_begin_0 = const()[name = tensor<string, []>("v_cache_21_begin_0"), val = tensor<int32, [5]>([10, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_21_end_0 = const()[name = tensor<string, []>("v_cache_21_end_0"), val = tensor<int32, [5]>([11, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_21_end_mask_0 = const()[name = tensor<string, []>("v_cache_21_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_21_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_21_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_21_cast_fp16 = slice_by_index(begin = v_cache_21_begin_0, end = v_cache_21_end_0, end_mask = v_cache_21_end_mask_0, squeeze_mask = v_cache_21_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_21_cast_fp16")];
            tensor<int32, []> var_1861 = const()[name = tensor<string, []>("op_1861"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_1860_promoted = const()[name = tensor<string, []>("op_1860_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_401_cast_fp16_to_fp32 = cast(dtype = x_401_cast_fp16_to_fp32_dtype_0, x = x_401_cast_fp16)[name = tensor<string, []>("cast_506")];
            tensor<fp32, [1, 1, 512]> var_1870 = pow(x = x_401_cast_fp16_to_fp32, y = var_1860_promoted)[name = tensor<string, []>("op_1870")];
            tensor<int32, [1]> var_81_axes_0 = const()[name = tensor<string, []>("var_81_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_81_keep_dims_0 = const()[name = tensor<string, []>("var_81_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_81 = reduce_mean(axes = var_81_axes_0, keep_dims = var_81_keep_dims_0, x = var_1870)[name = tensor<string, []>("var_81")];
            tensor<string, []> var_81_to_fp16_dtype_0_0 = const()[name = tensor<string, []>("var_81_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1874_to_fp16 = const()[name = tensor<string, []>("op_1874_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_81_to_fp16 = cast(dtype = var_81_to_fp16_dtype_0_0, x = var_81)[name = tensor<string, []>("cast_505")];
            tensor<fp16, [1, 1, 1]> var_1875_cast_fp16 = add(x = var_81_to_fp16, y = var_1874_to_fp16)[name = tensor<string, []>("op_1875_cast_fp16")];
            tensor<string, []> var_1875_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1875_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1876_epsilon_0 = const()[name = tensor<string, []>("op_1876_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1875_cast_fp16_to_fp32 = cast(dtype = var_1875_cast_fp16_to_fp32_dtype_0, x = var_1875_cast_fp16)[name = tensor<string, []>("cast_504")];
            tensor<fp32, [1, 1, 1]> var_1876 = rsqrt(epsilon = var_1876_epsilon_0, x = var_1875_cast_fp16_to_fp32)[name = tensor<string, []>("op_1876")];
            tensor<string, []> var_1876_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1876_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1876_to_fp16 = cast(dtype = var_1876_to_fp16_dtype_0, x = var_1876)[name = tensor<string, []>("cast_503")];
            tensor<fp16, [1, 1, 512]> x_407_cast_fp16 = mul(x = x_401_cast_fp16, y = var_1876_to_fp16)[name = tensor<string, []>("x_407_cast_fp16")];
            tensor<fp16, [512]> layers_10_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311620032)))];
            tensor<fp16, [1, 1, 512]> x_409_cast_fp16 = mul(x = layers_10_input_layernorm_weight_to_fp16, y = x_407_cast_fp16)[name = tensor<string, []>("x_409_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311621120)))];
            tensor<fp16, [1, 1, 1536]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = x_409_cast_fp16)[name = tensor<string, []>("linear_70_cast_fp16")];
            tensor<int32, [4]> var_1890 = const()[name = tensor<string, []>("op_1890"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_411_cast_fp16 = reshape(shape = var_1890, x = linear_70_cast_fp16)[name = tensor<string, []>("x_411_cast_fp16")];
            tensor<string, []> x_411_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_411_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1860_promoted_1 = const()[name = tensor<string, []>("op_1860_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_411_cast_fp16_to_fp32 = cast(dtype = x_411_cast_fp16_to_fp32_dtype_0, x = x_411_cast_fp16)[name = tensor<string, []>("cast_502")];
            tensor<fp32, [1, 1, 12, 128]> var_1894 = pow(x = x_411_cast_fp16_to_fp32, y = var_1860_promoted_1)[name = tensor<string, []>("op_1894")];
            tensor<int32, [1]> var_83_axes_0_0 = const()[name = tensor<string, []>("var_83_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_83_keep_dims_0 = const()[name = tensor<string, []>("var_83_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_83 = reduce_mean(axes = var_83_axes_0_0, keep_dims = var_83_keep_dims_0, x = var_1894)[name = tensor<string, []>("var_83")];
            tensor<string, []> var_83_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_83_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1898_to_fp16 = const()[name = tensor<string, []>("op_1898_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_83_to_fp16 = cast(dtype = var_83_to_fp16_dtype_0, x = var_83)[name = tensor<string, []>("cast_501")];
            tensor<fp16, [1, 1, 12, 1]> var_1899_cast_fp16 = add(x = var_83_to_fp16, y = var_1898_to_fp16)[name = tensor<string, []>("op_1899_cast_fp16")];
            tensor<string, []> var_1899_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1899_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1900_epsilon_0 = const()[name = tensor<string, []>("op_1900_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_1899_cast_fp16_to_fp32 = cast(dtype = var_1899_cast_fp16_to_fp32_dtype_0, x = var_1899_cast_fp16)[name = tensor<string, []>("cast_500")];
            tensor<fp32, [1, 1, 12, 1]> var_1900 = rsqrt(epsilon = var_1900_epsilon_0, x = var_1899_cast_fp16_to_fp32)[name = tensor<string, []>("op_1900")];
            tensor<string, []> var_1900_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1900_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_1900_to_fp16 = cast(dtype = var_1900_to_fp16_dtype_0, x = var_1900)[name = tensor<string, []>("cast_499")];
            tensor<fp16, [1, 1, 12, 128]> x_417_cast_fp16 = mul(x = x_411_cast_fp16, y = var_1900_to_fp16)[name = tensor<string, []>("x_417_cast_fp16")];
            tensor<fp16, [128]> layers_10_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313194048)))];
            tensor<fp16, [1, 1, 12, 128]> var_1902_cast_fp16 = mul(x = layers_10_self_attn_q_norm_weight_to_fp16, y = x_417_cast_fp16)[name = tensor<string, []>("op_1902_cast_fp16")];
            tensor<int32, [4]> q_41_perm_0 = const()[name = tensor<string, []>("q_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313194368)))];
            tensor<fp16, [1, 1, 512]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = x_409_cast_fp16)[name = tensor<string, []>("linear_71_cast_fp16")];
            tensor<int32, [4]> var_1906 = const()[name = tensor<string, []>("op_1906"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_419_cast_fp16 = reshape(shape = var_1906, x = linear_71_cast_fp16)[name = tensor<string, []>("x_419_cast_fp16")];
            tensor<string, []> x_419_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_419_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1860_promoted_2 = const()[name = tensor<string, []>("op_1860_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_419_cast_fp16_to_fp32 = cast(dtype = x_419_cast_fp16_to_fp32_dtype_0, x = x_419_cast_fp16)[name = tensor<string, []>("cast_498")];
            tensor<fp32, [1, 1, 4, 128]> var_1910 = pow(x = x_419_cast_fp16_to_fp32, y = var_1860_promoted_2)[name = tensor<string, []>("op_1910")];
            tensor<int32, [1]> var_85_axes_0 = const()[name = tensor<string, []>("var_85_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_85_keep_dims_0 = const()[name = tensor<string, []>("var_85_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_85 = reduce_mean(axes = var_85_axes_0, keep_dims = var_85_keep_dims_0, x = var_1910)[name = tensor<string, []>("var_85")];
            tensor<string, []> var_85_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_85_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1914_to_fp16 = const()[name = tensor<string, []>("op_1914_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_85_to_fp16 = cast(dtype = var_85_to_fp16_dtype_0, x = var_85)[name = tensor<string, []>("cast_497")];
            tensor<fp16, [1, 1, 4, 1]> var_1915_cast_fp16 = add(x = var_85_to_fp16, y = var_1914_to_fp16)[name = tensor<string, []>("op_1915_cast_fp16")];
            tensor<string, []> var_1915_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1915_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1916_epsilon_0 = const()[name = tensor<string, []>("op_1916_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_1915_cast_fp16_to_fp32 = cast(dtype = var_1915_cast_fp16_to_fp32_dtype_0, x = var_1915_cast_fp16)[name = tensor<string, []>("cast_496")];
            tensor<fp32, [1, 1, 4, 1]> var_1916 = rsqrt(epsilon = var_1916_epsilon_0, x = var_1915_cast_fp16_to_fp32)[name = tensor<string, []>("op_1916")];
            tensor<string, []> var_1916_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1916_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_1916_to_fp16 = cast(dtype = var_1916_to_fp16_dtype_0, x = var_1916)[name = tensor<string, []>("cast_495")];
            tensor<fp16, [1, 1, 4, 128]> x_425_cast_fp16 = mul(x = x_419_cast_fp16, y = var_1916_to_fp16)[name = tensor<string, []>("x_425_cast_fp16")];
            tensor<fp16, [128]> layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313718720)))];
            tensor<fp16, [1, 1, 4, 128]> var_1918_cast_fp16 = mul(x = layers_10_self_attn_k_norm_weight_to_fp16, y = x_425_cast_fp16)[name = tensor<string, []>("op_1918_cast_fp16")];
            tensor<int32, [4]> k_41_perm_0 = const()[name = tensor<string, []>("k_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(313719040)))];
            tensor<fp16, [1, 1, 512]> linear_72_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = x_409_cast_fp16)[name = tensor<string, []>("linear_72_cast_fp16")];
            tensor<int32, [4]> var_1922 = const()[name = tensor<string, []>("op_1922"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_1923_cast_fp16 = reshape(shape = var_1922, x = linear_72_cast_fp16)[name = tensor<string, []>("op_1923_cast_fp16")];
            tensor<int32, [4]> v_21_perm_0 = const()[name = tensor<string, []>("v_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_41_cast_fp16 = transpose(perm = q_41_perm_0, x = var_1902_cast_fp16)[name = tensor<string, []>("transpose_71")];
            tensor<fp16, [1, 12, 1, 128]> var_1927_cast_fp16 = mul(x = q_41_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1927_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = tensor<string, []>("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = tensor<string, []>("x1_41_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = tensor<string, []>("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_41_cast_fp16)[name = tensor<string, []>("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = tensor<string, []>("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = tensor<string, []>("x2_41_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = tensor<string, []>("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_41_cast_fp16)[name = tensor<string, []>("x2_41_cast_fp16")];
            tensor<fp16, []> const_21_promoted_to_fp16 = const()[name = tensor<string, []>("const_21_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_1930_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_21_promoted_to_fp16)[name = tensor<string, []>("op_1930_cast_fp16")];
            tensor<bool, []> var_1932_interleave_0 = const()[name = tensor<string, []>("op_1932_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_1932_cast_fp16 = concat(axis = var_1861, interleave = var_1932_interleave_0, values = (var_1930_cast_fp16, x1_41_cast_fp16))[name = tensor<string, []>("op_1932_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_1933_cast_fp16 = mul(x = var_1932_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1933_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_43_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1933_cast_fp16)[name = tensor<string, []>("q_43_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_41_cast_fp16 = transpose(perm = k_41_perm_0, x = var_1918_cast_fp16)[name = tensor<string, []>("transpose_70")];
            tensor<fp16, [1, 4, 1, 128]> var_1935_cast_fp16 = mul(x = k_41_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_1935_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = tensor<string, []>("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = tensor<string, []>("x1_43_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = tensor<string, []>("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_41_cast_fp16)[name = tensor<string, []>("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = tensor<string, []>("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = tensor<string, []>("x2_43_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = tensor<string, []>("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_41_cast_fp16)[name = tensor<string, []>("x2_43_cast_fp16")];
            tensor<fp16, []> const_22_promoted_to_fp16 = const()[name = tensor<string, []>("const_22_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_1938_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_22_promoted_to_fp16)[name = tensor<string, []>("op_1938_cast_fp16")];
            tensor<bool, []> var_1940_interleave_0 = const()[name = tensor<string, []>("op_1940_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_1940_cast_fp16 = concat(axis = var_1861, interleave = var_1940_interleave_0, values = (var_1938_cast_fp16, x1_43_cast_fp16))[name = tensor<string, []>("op_1940_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_1941_cast_fp16 = mul(x = var_1940_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_1941_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_43_cast_fp16 = add(x = var_1935_cast_fp16, y = var_1941_cast_fp16)[name = tensor<string, []>("k_43_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1944_cast_fp16 = mul(x = k_cache_21_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1944_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1945_cast_fp16 = mul(x = k_43_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1945_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_21_cast_fp16 = add(x = var_1944_cast_fp16, y = var_1945_cast_fp16)[name = tensor<string, []>("k_full_21_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_1948_cast_fp16 = mul(x = v_cache_21_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_1948_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_1923_cast_fp16)[name = tensor<string, []>("transpose_69")];
            tensor<fp16, [1, 4, 2048, 128]> var_1949_cast_fp16 = mul(x = v_21_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_1949_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_21_cast_fp16 = add(x = var_1948_cast_fp16, y = var_1949_cast_fp16)[name = tensor<string, []>("v_full_21_cast_fp16")];
            tensor<int32, [1]> var_1951_axes_0 = const()[name = tensor<string, []>("op_1951_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1951_cast_fp16 = expand_dims(axes = var_1951_axes_0, x = k_full_21_cast_fp16)[name = tensor<string, []>("op_1951_cast_fp16")];
            tensor<int32, [5]> var_1953_reps_0 = const()[name = tensor<string, []>("op_1953_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1953_cast_fp16 = tile(reps = var_1953_reps_0, x = var_1951_cast_fp16)[name = tensor<string, []>("op_1953_cast_fp16")];
            tensor<int32, [4]> var_1954 = const()[name = tensor<string, []>("op_1954"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_21_cast_fp16 = reshape(shape = var_1954, x = var_1953_cast_fp16)[name = tensor<string, []>("k_rep_21_cast_fp16")];
            tensor<int32, [1]> var_1956_axes_0 = const()[name = tensor<string, []>("op_1956_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_1956_cast_fp16 = expand_dims(axes = var_1956_axes_0, x = v_full_21_cast_fp16)[name = tensor<string, []>("op_1956_cast_fp16")];
            tensor<int32, [5]> var_1958_reps_0 = const()[name = tensor<string, []>("op_1958_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_1958_cast_fp16 = tile(reps = var_1958_reps_0, x = var_1956_cast_fp16)[name = tensor<string, []>("op_1958_cast_fp16")];
            tensor<int32, [4]> var_1959 = const()[name = tensor<string, []>("op_1959"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_21_cast_fp16 = reshape(shape = var_1959, x = var_1958_cast_fp16)[name = tensor<string, []>("v_rep_21_cast_fp16")];
            tensor<bool, []> var_1962_transpose_x_1 = const()[name = tensor<string, []>("op_1962_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_1962_transpose_y_1 = const()[name = tensor<string, []>("op_1962_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_1962_cast_fp16 = matmul(transpose_x = var_1962_transpose_x_1, transpose_y = var_1962_transpose_y_1, x = q_43_cast_fp16, y = k_rep_21_cast_fp16)[name = tensor<string, []>("op_1962_cast_fp16")];
            tensor<fp16, []> var_1963_to_fp16 = const()[name = tensor<string, []>("op_1963_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_41_cast_fp16 = mul(x = var_1962_cast_fp16, y = var_1963_to_fp16)[name = tensor<string, []>("attn_41_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_43_cast_fp16 = add(x = attn_41_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_43_cast_fp16")];
            tensor<string, []> input_43_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_43_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_43_cast_fp16_to_fp32 = cast(dtype = input_43_cast_fp16_to_fp32_dtype_0, x = input_43_cast_fp16)[name = tensor<string, []>("cast_494")];
            tensor<fp32, [1, 12, 1, 2048]> attn_43 = softmax(axis = var_1861, x = input_43_cast_fp16_to_fp32)[name = tensor<string, []>("attn_43")];
            tensor<bool, []> out_21_transpose_x_0 = const()[name = tensor<string, []>("out_21_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_21_transpose_y_0 = const()[name = tensor<string, []>("out_21_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_43_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_43_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_43_to_fp16 = cast(dtype = attn_43_to_fp16_dtype_0, x = attn_43)[name = tensor<string, []>("cast_493")];
            tensor<fp16, [1, 12, 1, 128]> out_21_cast_fp16 = matmul(transpose_x = out_21_transpose_x_0, transpose_y = out_21_transpose_y_0, x = attn_43_to_fp16, y = v_rep_21_cast_fp16)[name = tensor<string, []>("out_21_cast_fp16")];
            tensor<int32, [4]> var_1968_perm_0 = const()[name = tensor<string, []>("op_1968_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1970 = const()[name = tensor<string, []>("op_1970"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_1968_cast_fp16 = transpose(perm = var_1968_perm_0, x = out_21_cast_fp16)[name = tensor<string, []>("transpose_68")];
            tensor<fp16, [1, 1, 1536]> x_427_cast_fp16 = reshape(shape = var_1970, x = var_1968_cast_fp16)[name = tensor<string, []>("x_427_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(314243392)))];
            tensor<fp16, [1, 1, 512]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = x_427_cast_fp16)[name = tensor<string, []>("linear_73_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_429_cast_fp16 = add(x = x_401_cast_fp16, y = linear_73_cast_fp16)[name = tensor<string, []>("x_429_cast_fp16")];
            tensor<string, []> x_429_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_429_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1860_promoted_3 = const()[name = tensor<string, []>("op_1860_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_429_cast_fp16_to_fp32 = cast(dtype = x_429_cast_fp16_to_fp32_dtype_0, x = x_429_cast_fp16)[name = tensor<string, []>("cast_492")];
            tensor<fp32, [1, 1, 512]> var_1981 = pow(x = x_429_cast_fp16_to_fp32, y = var_1860_promoted_3)[name = tensor<string, []>("op_1981")];
            tensor<int32, [1]> var_87_axes_0 = const()[name = tensor<string, []>("var_87_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_87_keep_dims_0 = const()[name = tensor<string, []>("var_87_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_87 = reduce_mean(axes = var_87_axes_0, keep_dims = var_87_keep_dims_0, x = var_1981)[name = tensor<string, []>("var_87")];
            tensor<string, []> var_87_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_87_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_1985_to_fp16 = const()[name = tensor<string, []>("op_1985_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_87_to_fp16 = cast(dtype = var_87_to_fp16_dtype_0, x = var_87)[name = tensor<string, []>("cast_491")];
            tensor<fp16, [1, 1, 1]> var_1986_cast_fp16 = add(x = var_87_to_fp16, y = var_1985_to_fp16)[name = tensor<string, []>("op_1986_cast_fp16")];
            tensor<string, []> var_1986_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_1986_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_1987_epsilon_0 = const()[name = tensor<string, []>("op_1987_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_1986_cast_fp16_to_fp32 = cast(dtype = var_1986_cast_fp16_to_fp32_dtype_0, x = var_1986_cast_fp16)[name = tensor<string, []>("cast_490")];
            tensor<fp32, [1, 1, 1]> var_1987 = rsqrt(epsilon = var_1987_epsilon_0, x = var_1986_cast_fp16_to_fp32)[name = tensor<string, []>("op_1987")];
            tensor<string, []> var_1987_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_1987_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_1987_to_fp16 = cast(dtype = var_1987_to_fp16_dtype_0, x = var_1987)[name = tensor<string, []>("cast_489")];
            tensor<fp16, [1, 1, 512]> x_435_cast_fp16 = mul(x = x_429_cast_fp16, y = var_1987_to_fp16)[name = tensor<string, []>("x_435_cast_fp16")];
            tensor<fp16, [512]> layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315816320)))];
            tensor<fp16, [1, 1, 512]> x_437_cast_fp16 = mul(x = layers_10_post_attention_layernorm_weight_to_fp16, y = x_435_cast_fp16)[name = tensor<string, []>("x_437_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_10_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(315817408)))];
            tensor<fp16, [1, 1, 1536]> linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_mlp_gate_proj_weight_to_fp16, x = x_437_cast_fp16)[name = tensor<string, []>("linear_74_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_1998_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor<string, []>("op_1998_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_10_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(317390336)))];
            tensor<fp16, [1, 1, 1536]> linear_75_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_mlp_up_proj_weight_to_fp16, x = x_437_cast_fp16)[name = tensor<string, []>("linear_75_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_439_cast_fp16 = mul(x = var_1998_cast_fp16, y = linear_75_cast_fp16)[name = tensor<string, []>("x_439_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_10_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_10_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318963264)))];
            tensor<fp16, [1, 1, 512]> linear_76_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_10_mlp_down_proj_weight_to_fp16, x = x_439_cast_fp16)[name = tensor<string, []>("linear_76_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_441_cast_fp16 = add(x = x_429_cast_fp16, y = linear_76_cast_fp16)[name = tensor<string, []>("x_441_cast_fp16")];
            tensor<string, []> x_441_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_441_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_23_begin_0 = const()[name = tensor<string, []>("k_cache_23_begin_0"), val = tensor<int32, [5]>([11, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_23_end_0 = const()[name = tensor<string, []>("k_cache_23_end_0"), val = tensor<int32, [5]>([12, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_23_end_mask_0 = const()[name = tensor<string, []>("k_cache_23_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_23_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_23_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_23_cast_fp16 = slice_by_index(begin = k_cache_23_begin_0, end = k_cache_23_end_0, end_mask = k_cache_23_end_mask_0, squeeze_mask = k_cache_23_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_23_cast_fp16")];
            tensor<int32, [5]> v_cache_23_begin_0 = const()[name = tensor<string, []>("v_cache_23_begin_0"), val = tensor<int32, [5]>([11, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_23_end_0 = const()[name = tensor<string, []>("v_cache_23_end_0"), val = tensor<int32, [5]>([12, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_23_end_mask_0 = const()[name = tensor<string, []>("v_cache_23_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_23_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_23_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_23_cast_fp16 = slice_by_index(begin = v_cache_23_begin_0, end = v_cache_23_end_0, end_mask = v_cache_23_end_mask_0, squeeze_mask = v_cache_23_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_23_cast_fp16")];
            tensor<int32, []> var_2030 = const()[name = tensor<string, []>("op_2030"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_2029_promoted = const()[name = tensor<string, []>("op_2029_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_441_cast_fp16_to_fp32 = cast(dtype = x_441_cast_fp16_to_fp32_dtype_0, x = x_441_cast_fp16)[name = tensor<string, []>("cast_488")];
            tensor<fp32, [1, 1, 512]> var_2039 = pow(x = x_441_cast_fp16_to_fp32, y = var_2029_promoted)[name = tensor<string, []>("op_2039")];
            tensor<int32, [1]> var_89_axes_0 = const()[name = tensor<string, []>("var_89_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_89_keep_dims_0 = const()[name = tensor<string, []>("var_89_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_89 = reduce_mean(axes = var_89_axes_0, keep_dims = var_89_keep_dims_0, x = var_2039)[name = tensor<string, []>("var_89")];
            tensor<string, []> var_89_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_89_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2043_to_fp16 = const()[name = tensor<string, []>("op_2043_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_89_to_fp16 = cast(dtype = var_89_to_fp16_dtype_0, x = var_89)[name = tensor<string, []>("cast_487")];
            tensor<fp16, [1, 1, 1]> var_2044_cast_fp16 = add(x = var_89_to_fp16, y = var_2043_to_fp16)[name = tensor<string, []>("op_2044_cast_fp16")];
            tensor<string, []> var_2044_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2044_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2045_epsilon_0 = const()[name = tensor<string, []>("op_2045_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2044_cast_fp16_to_fp32 = cast(dtype = var_2044_cast_fp16_to_fp32_dtype_0, x = var_2044_cast_fp16)[name = tensor<string, []>("cast_486")];
            tensor<fp32, [1, 1, 1]> var_2045 = rsqrt(epsilon = var_2045_epsilon_0, x = var_2044_cast_fp16_to_fp32)[name = tensor<string, []>("op_2045")];
            tensor<string, []> var_2045_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2045_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2045_to_fp16 = cast(dtype = var_2045_to_fp16_dtype_0, x = var_2045)[name = tensor<string, []>("cast_485")];
            tensor<fp16, [1, 1, 512]> x_447_cast_fp16 = mul(x = x_441_cast_fp16, y = var_2045_to_fp16)[name = tensor<string, []>("x_447_cast_fp16")];
            tensor<fp16, [512]> layers_11_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320536192)))];
            tensor<fp16, [1, 1, 512]> x_449_cast_fp16 = mul(x = layers_11_input_layernorm_weight_to_fp16, y = x_447_cast_fp16)[name = tensor<string, []>("x_449_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320537280)))];
            tensor<fp16, [1, 1, 1536]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = x_449_cast_fp16)[name = tensor<string, []>("linear_77_cast_fp16")];
            tensor<int32, [4]> var_2059 = const()[name = tensor<string, []>("op_2059"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_451_cast_fp16 = reshape(shape = var_2059, x = linear_77_cast_fp16)[name = tensor<string, []>("x_451_cast_fp16")];
            tensor<string, []> x_451_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_451_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2029_promoted_1 = const()[name = tensor<string, []>("op_2029_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_451_cast_fp16_to_fp32 = cast(dtype = x_451_cast_fp16_to_fp32_dtype_0, x = x_451_cast_fp16)[name = tensor<string, []>("cast_484")];
            tensor<fp32, [1, 1, 12, 128]> var_2063 = pow(x = x_451_cast_fp16_to_fp32, y = var_2029_promoted_1)[name = tensor<string, []>("op_2063")];
            tensor<int32, [1]> var_91_axes_0 = const()[name = tensor<string, []>("var_91_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_91_keep_dims_0 = const()[name = tensor<string, []>("var_91_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_91_0 = reduce_mean(axes = var_91_axes_0, keep_dims = var_91_keep_dims_0, x = var_2063)[name = tensor<string, []>("var_91")];
            tensor<string, []> var_91_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_91_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2067_to_fp16 = const()[name = tensor<string, []>("op_2067_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_91_to_fp16 = cast(dtype = var_91_to_fp16_dtype_0, x = var_91_0)[name = tensor<string, []>("cast_483")];
            tensor<fp16, [1, 1, 12, 1]> var_2068_cast_fp16 = add(x = var_91_to_fp16, y = var_2067_to_fp16)[name = tensor<string, []>("op_2068_cast_fp16")];
            tensor<string, []> var_2068_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2068_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2069_epsilon_0 = const()[name = tensor<string, []>("op_2069_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_2068_cast_fp16_to_fp32 = cast(dtype = var_2068_cast_fp16_to_fp32_dtype_0, x = var_2068_cast_fp16)[name = tensor<string, []>("cast_482")];
            tensor<fp32, [1, 1, 12, 1]> var_2069 = rsqrt(epsilon = var_2069_epsilon_0, x = var_2068_cast_fp16_to_fp32)[name = tensor<string, []>("op_2069")];
            tensor<string, []> var_2069_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2069_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_2069_to_fp16 = cast(dtype = var_2069_to_fp16_dtype_0, x = var_2069)[name = tensor<string, []>("cast_481")];
            tensor<fp16, [1, 1, 12, 128]> x_457_cast_fp16 = mul(x = x_451_cast_fp16, y = var_2069_to_fp16)[name = tensor<string, []>("x_457_cast_fp16")];
            tensor<fp16, [128]> layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322110208)))];
            tensor<fp16, [1, 1, 12, 128]> var_2071_cast_fp16 = mul(x = layers_11_self_attn_q_norm_weight_to_fp16, y = x_457_cast_fp16)[name = tensor<string, []>("op_2071_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = tensor<string, []>("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322110528)))];
            tensor<fp16, [1, 1, 512]> linear_78_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = x_449_cast_fp16)[name = tensor<string, []>("linear_78_cast_fp16")];
            tensor<int32, [4]> var_2075 = const()[name = tensor<string, []>("op_2075"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_459_cast_fp16 = reshape(shape = var_2075, x = linear_78_cast_fp16)[name = tensor<string, []>("x_459_cast_fp16")];
            tensor<string, []> x_459_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_459_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2029_promoted_2 = const()[name = tensor<string, []>("op_2029_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_459_cast_fp16_to_fp32 = cast(dtype = x_459_cast_fp16_to_fp32_dtype_0, x = x_459_cast_fp16)[name = tensor<string, []>("cast_480")];
            tensor<fp32, [1, 1, 4, 128]> var_2079 = pow(x = x_459_cast_fp16_to_fp32, y = var_2029_promoted_2)[name = tensor<string, []>("op_2079")];
            tensor<int32, [1]> var_93_axes_0 = const()[name = tensor<string, []>("var_93_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_93_keep_dims_0 = const()[name = tensor<string, []>("var_93_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_93 = reduce_mean(axes = var_93_axes_0, keep_dims = var_93_keep_dims_0, x = var_2079)[name = tensor<string, []>("var_93")];
            tensor<string, []> var_93_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_93_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2083_to_fp16 = const()[name = tensor<string, []>("op_2083_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_93_to_fp16 = cast(dtype = var_93_to_fp16_dtype_0, x = var_93)[name = tensor<string, []>("cast_479")];
            tensor<fp16, [1, 1, 4, 1]> var_2084_cast_fp16 = add(x = var_93_to_fp16, y = var_2083_to_fp16)[name = tensor<string, []>("op_2084_cast_fp16")];
            tensor<string, []> var_2084_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2084_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2085_epsilon_0 = const()[name = tensor<string, []>("op_2085_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_2084_cast_fp16_to_fp32 = cast(dtype = var_2084_cast_fp16_to_fp32_dtype_0, x = var_2084_cast_fp16)[name = tensor<string, []>("cast_478")];
            tensor<fp32, [1, 1, 4, 1]> var_2085 = rsqrt(epsilon = var_2085_epsilon_0, x = var_2084_cast_fp16_to_fp32)[name = tensor<string, []>("op_2085")];
            tensor<string, []> var_2085_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2085_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_2085_to_fp16 = cast(dtype = var_2085_to_fp16_dtype_0, x = var_2085)[name = tensor<string, []>("cast_477")];
            tensor<fp16, [1, 1, 4, 128]> x_465_cast_fp16 = mul(x = x_459_cast_fp16, y = var_2085_to_fp16)[name = tensor<string, []>("x_465_cast_fp16")];
            tensor<fp16, [128]> layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322634880)))];
            tensor<fp16, [1, 1, 4, 128]> var_2087_cast_fp16 = mul(x = layers_11_self_attn_k_norm_weight_to_fp16, y = x_465_cast_fp16)[name = tensor<string, []>("op_2087_cast_fp16")];
            tensor<int32, [4]> k_45_perm_0 = const()[name = tensor<string, []>("k_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(322635200)))];
            tensor<fp16, [1, 1, 512]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = x_449_cast_fp16)[name = tensor<string, []>("linear_79_cast_fp16")];
            tensor<int32, [4]> var_2091 = const()[name = tensor<string, []>("op_2091"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_2092_cast_fp16 = reshape(shape = var_2091, x = linear_79_cast_fp16)[name = tensor<string, []>("op_2092_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = tensor<string, []>("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = var_2071_cast_fp16)[name = tensor<string, []>("transpose_67")];
            tensor<fp16, [1, 12, 1, 128]> var_2096_cast_fp16 = mul(x = q_45_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2096_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = tensor<string, []>("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = tensor<string, []>("x1_45_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = tensor<string, []>("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_45_cast_fp16)[name = tensor<string, []>("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = tensor<string, []>("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = tensor<string, []>("x2_45_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = tensor<string, []>("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_45_cast_fp16)[name = tensor<string, []>("x2_45_cast_fp16")];
            tensor<fp16, []> const_23_promoted_to_fp16 = const()[name = tensor<string, []>("const_23_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_2099_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_23_promoted_to_fp16)[name = tensor<string, []>("op_2099_cast_fp16")];
            tensor<bool, []> var_2101_interleave_0 = const()[name = tensor<string, []>("op_2101_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_2101_cast_fp16 = concat(axis = var_2030, interleave = var_2101_interleave_0, values = (var_2099_cast_fp16, x1_45_cast_fp16))[name = tensor<string, []>("op_2101_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_2102_cast_fp16 = mul(x = var_2101_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2102_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_47_cast_fp16 = add(x = var_2096_cast_fp16, y = var_2102_cast_fp16)[name = tensor<string, []>("q_47_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_45_cast_fp16 = transpose(perm = k_45_perm_0, x = var_2087_cast_fp16)[name = tensor<string, []>("transpose_66")];
            tensor<fp16, [1, 4, 1, 128]> var_2104_cast_fp16 = mul(x = k_45_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2104_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = tensor<string, []>("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = tensor<string, []>("x1_47_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = tensor<string, []>("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_45_cast_fp16)[name = tensor<string, []>("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = tensor<string, []>("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = tensor<string, []>("x2_47_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = tensor<string, []>("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_45_cast_fp16)[name = tensor<string, []>("x2_47_cast_fp16")];
            tensor<fp16, []> const_24_promoted_to_fp16 = const()[name = tensor<string, []>("const_24_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_2107_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_24_promoted_to_fp16)[name = tensor<string, []>("op_2107_cast_fp16")];
            tensor<bool, []> var_2109_interleave_0 = const()[name = tensor<string, []>("op_2109_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_2109_cast_fp16 = concat(axis = var_2030, interleave = var_2109_interleave_0, values = (var_2107_cast_fp16, x1_47_cast_fp16))[name = tensor<string, []>("op_2109_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_2110_cast_fp16 = mul(x = var_2109_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2110_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_47_cast_fp16 = add(x = var_2104_cast_fp16, y = var_2110_cast_fp16)[name = tensor<string, []>("k_47_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2113_cast_fp16 = mul(x = k_cache_23_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2113_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2114_cast_fp16 = mul(x = k_47_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2114_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_23_cast_fp16 = add(x = var_2113_cast_fp16, y = var_2114_cast_fp16)[name = tensor<string, []>("k_full_23_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2117_cast_fp16 = mul(x = v_cache_23_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2117_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_2092_cast_fp16)[name = tensor<string, []>("transpose_65")];
            tensor<fp16, [1, 4, 2048, 128]> var_2118_cast_fp16 = mul(x = v_23_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2118_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_23_cast_fp16 = add(x = var_2117_cast_fp16, y = var_2118_cast_fp16)[name = tensor<string, []>("v_full_23_cast_fp16")];
            tensor<int32, [1]> var_2120_axes_0 = const()[name = tensor<string, []>("op_2120_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2120_cast_fp16 = expand_dims(axes = var_2120_axes_0, x = k_full_23_cast_fp16)[name = tensor<string, []>("op_2120_cast_fp16")];
            tensor<int32, [5]> var_2122_reps_0 = const()[name = tensor<string, []>("op_2122_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2122_cast_fp16 = tile(reps = var_2122_reps_0, x = var_2120_cast_fp16)[name = tensor<string, []>("op_2122_cast_fp16")];
            tensor<int32, [4]> var_2123 = const()[name = tensor<string, []>("op_2123"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_23_cast_fp16 = reshape(shape = var_2123, x = var_2122_cast_fp16)[name = tensor<string, []>("k_rep_23_cast_fp16")];
            tensor<int32, [1]> var_2125_axes_0 = const()[name = tensor<string, []>("op_2125_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2125_cast_fp16 = expand_dims(axes = var_2125_axes_0, x = v_full_23_cast_fp16)[name = tensor<string, []>("op_2125_cast_fp16")];
            tensor<int32, [5]> var_2127_reps_0 = const()[name = tensor<string, []>("op_2127_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2127_cast_fp16 = tile(reps = var_2127_reps_0, x = var_2125_cast_fp16)[name = tensor<string, []>("op_2127_cast_fp16")];
            tensor<int32, [4]> var_2128 = const()[name = tensor<string, []>("op_2128"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_23_cast_fp16 = reshape(shape = var_2128, x = var_2127_cast_fp16)[name = tensor<string, []>("v_rep_23_cast_fp16")];
            tensor<bool, []> var_2131_transpose_x_1 = const()[name = tensor<string, []>("op_2131_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2131_transpose_y_1 = const()[name = tensor<string, []>("op_2131_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_2131_cast_fp16 = matmul(transpose_x = var_2131_transpose_x_1, transpose_y = var_2131_transpose_y_1, x = q_47_cast_fp16, y = k_rep_23_cast_fp16)[name = tensor<string, []>("op_2131_cast_fp16")];
            tensor<fp16, []> var_2132_to_fp16 = const()[name = tensor<string, []>("op_2132_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_45_cast_fp16 = mul(x = var_2131_cast_fp16, y = var_2132_to_fp16)[name = tensor<string, []>("attn_45_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_47_cast_fp16 = add(x = attn_45_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_47_cast_fp16")];
            tensor<string, []> input_47_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_47_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_47_cast_fp16_to_fp32 = cast(dtype = input_47_cast_fp16_to_fp32_dtype_0, x = input_47_cast_fp16)[name = tensor<string, []>("cast_476")];
            tensor<fp32, [1, 12, 1, 2048]> attn_47 = softmax(axis = var_2030, x = input_47_cast_fp16_to_fp32)[name = tensor<string, []>("attn_47")];
            tensor<bool, []> out_23_transpose_x_0 = const()[name = tensor<string, []>("out_23_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_23_transpose_y_0 = const()[name = tensor<string, []>("out_23_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_47_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_47_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_47_to_fp16 = cast(dtype = attn_47_to_fp16_dtype_0, x = attn_47)[name = tensor<string, []>("cast_475")];
            tensor<fp16, [1, 12, 1, 128]> out_23_cast_fp16 = matmul(transpose_x = out_23_transpose_x_0, transpose_y = out_23_transpose_y_0, x = attn_47_to_fp16, y = v_rep_23_cast_fp16)[name = tensor<string, []>("out_23_cast_fp16")];
            tensor<int32, [4]> var_2137_perm_0 = const()[name = tensor<string, []>("op_2137_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2139 = const()[name = tensor<string, []>("op_2139"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_2137_cast_fp16 = transpose(perm = var_2137_perm_0, x = out_23_cast_fp16)[name = tensor<string, []>("transpose_64")];
            tensor<fp16, [1, 1, 1536]> x_467_cast_fp16 = reshape(shape = var_2139, x = var_2137_cast_fp16)[name = tensor<string, []>("x_467_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(323159552)))];
            tensor<fp16, [1, 1, 512]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = x_467_cast_fp16)[name = tensor<string, []>("linear_80_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_469_cast_fp16 = add(x = x_441_cast_fp16, y = linear_80_cast_fp16)[name = tensor<string, []>("x_469_cast_fp16")];
            tensor<string, []> x_469_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_469_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2029_promoted_3 = const()[name = tensor<string, []>("op_2029_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_469_cast_fp16_to_fp32 = cast(dtype = x_469_cast_fp16_to_fp32_dtype_0, x = x_469_cast_fp16)[name = tensor<string, []>("cast_474")];
            tensor<fp32, [1, 1, 512]> var_2150 = pow(x = x_469_cast_fp16_to_fp32, y = var_2029_promoted_3)[name = tensor<string, []>("op_2150")];
            tensor<int32, [1]> var_95_axes_0 = const()[name = tensor<string, []>("var_95_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_95_keep_dims_0 = const()[name = tensor<string, []>("var_95_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_95 = reduce_mean(axes = var_95_axes_0, keep_dims = var_95_keep_dims_0, x = var_2150)[name = tensor<string, []>("var_95")];
            tensor<string, []> var_95_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_95_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2154_to_fp16 = const()[name = tensor<string, []>("op_2154_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_95_to_fp16 = cast(dtype = var_95_to_fp16_dtype_0, x = var_95)[name = tensor<string, []>("cast_473")];
            tensor<fp16, [1, 1, 1]> var_2155_cast_fp16 = add(x = var_95_to_fp16, y = var_2154_to_fp16)[name = tensor<string, []>("op_2155_cast_fp16")];
            tensor<string, []> var_2155_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2155_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2156_epsilon_0 = const()[name = tensor<string, []>("op_2156_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2155_cast_fp16_to_fp32 = cast(dtype = var_2155_cast_fp16_to_fp32_dtype_0, x = var_2155_cast_fp16)[name = tensor<string, []>("cast_472")];
            tensor<fp32, [1, 1, 1]> var_2156 = rsqrt(epsilon = var_2156_epsilon_0, x = var_2155_cast_fp16_to_fp32)[name = tensor<string, []>("op_2156")];
            tensor<string, []> var_2156_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2156_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2156_to_fp16 = cast(dtype = var_2156_to_fp16_dtype_0, x = var_2156)[name = tensor<string, []>("cast_471")];
            tensor<fp16, [1, 1, 512]> x_475_cast_fp16 = mul(x = x_469_cast_fp16, y = var_2156_to_fp16)[name = tensor<string, []>("x_475_cast_fp16")];
            tensor<fp16, [512]> layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324732480)))];
            tensor<fp16, [1, 1, 512]> x_477_cast_fp16 = mul(x = layers_11_post_attention_layernorm_weight_to_fp16, y = x_475_cast_fp16)[name = tensor<string, []>("x_477_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_11_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324733568)))];
            tensor<fp16, [1, 1, 1536]> linear_81_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_mlp_gate_proj_weight_to_fp16, x = x_477_cast_fp16)[name = tensor<string, []>("linear_81_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_2167_cast_fp16 = silu(x = linear_81_cast_fp16)[name = tensor<string, []>("op_2167_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_11_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(326306496)))];
            tensor<fp16, [1, 1, 1536]> linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_mlp_up_proj_weight_to_fp16, x = x_477_cast_fp16)[name = tensor<string, []>("linear_82_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_479_cast_fp16 = mul(x = var_2167_cast_fp16, y = linear_82_cast_fp16)[name = tensor<string, []>("x_479_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_11_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_11_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(327879424)))];
            tensor<fp16, [1, 1, 512]> linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_11_mlp_down_proj_weight_to_fp16, x = x_479_cast_fp16)[name = tensor<string, []>("linear_83_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_481_cast_fp16 = add(x = x_469_cast_fp16, y = linear_83_cast_fp16)[name = tensor<string, []>("x_481_cast_fp16")];
            tensor<string, []> x_481_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_481_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_25_begin_0 = const()[name = tensor<string, []>("k_cache_25_begin_0"), val = tensor<int32, [5]>([12, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_25_end_0 = const()[name = tensor<string, []>("k_cache_25_end_0"), val = tensor<int32, [5]>([13, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_25_end_mask_0 = const()[name = tensor<string, []>("k_cache_25_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_25_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_25_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_25_cast_fp16 = slice_by_index(begin = k_cache_25_begin_0, end = k_cache_25_end_0, end_mask = k_cache_25_end_mask_0, squeeze_mask = k_cache_25_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_25_cast_fp16")];
            tensor<int32, [5]> v_cache_25_begin_0 = const()[name = tensor<string, []>("v_cache_25_begin_0"), val = tensor<int32, [5]>([12, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_25_end_0 = const()[name = tensor<string, []>("v_cache_25_end_0"), val = tensor<int32, [5]>([13, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_25_end_mask_0 = const()[name = tensor<string, []>("v_cache_25_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_25_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_25_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_25_cast_fp16 = slice_by_index(begin = v_cache_25_begin_0, end = v_cache_25_end_0, end_mask = v_cache_25_end_mask_0, squeeze_mask = v_cache_25_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_25_cast_fp16")];
            tensor<int32, []> var_2199 = const()[name = tensor<string, []>("op_2199"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_2198_promoted = const()[name = tensor<string, []>("op_2198_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_481_cast_fp16_to_fp32 = cast(dtype = x_481_cast_fp16_to_fp32_dtype_0, x = x_481_cast_fp16)[name = tensor<string, []>("cast_470")];
            tensor<fp32, [1, 1, 512]> var_2208 = pow(x = x_481_cast_fp16_to_fp32, y = var_2198_promoted)[name = tensor<string, []>("op_2208")];
            tensor<int32, [1]> var_97_axes_0 = const()[name = tensor<string, []>("var_97_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_97_keep_dims_0 = const()[name = tensor<string, []>("var_97_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_97 = reduce_mean(axes = var_97_axes_0, keep_dims = var_97_keep_dims_0, x = var_2208)[name = tensor<string, []>("var_97")];
            tensor<string, []> var_97_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_97_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2212_to_fp16 = const()[name = tensor<string, []>("op_2212_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_97_to_fp16 = cast(dtype = var_97_to_fp16_dtype_0, x = var_97)[name = tensor<string, []>("cast_469")];
            tensor<fp16, [1, 1, 1]> var_2213_cast_fp16 = add(x = var_97_to_fp16, y = var_2212_to_fp16)[name = tensor<string, []>("op_2213_cast_fp16")];
            tensor<string, []> var_2213_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2213_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2214_epsilon_0 = const()[name = tensor<string, []>("op_2214_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2213_cast_fp16_to_fp32 = cast(dtype = var_2213_cast_fp16_to_fp32_dtype_0, x = var_2213_cast_fp16)[name = tensor<string, []>("cast_468")];
            tensor<fp32, [1, 1, 1]> var_2214 = rsqrt(epsilon = var_2214_epsilon_0, x = var_2213_cast_fp16_to_fp32)[name = tensor<string, []>("op_2214")];
            tensor<string, []> var_2214_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2214_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2214_to_fp16 = cast(dtype = var_2214_to_fp16_dtype_0, x = var_2214)[name = tensor<string, []>("cast_467")];
            tensor<fp16, [1, 1, 512]> x_487_cast_fp16 = mul(x = x_481_cast_fp16, y = var_2214_to_fp16)[name = tensor<string, []>("x_487_cast_fp16")];
            tensor<fp16, [512]> layers_12_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329452352)))];
            tensor<fp16, [1, 1, 512]> x_489_cast_fp16 = mul(x = layers_12_input_layernorm_weight_to_fp16, y = x_487_cast_fp16)[name = tensor<string, []>("x_489_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329453440)))];
            tensor<fp16, [1, 1, 1536]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = x_489_cast_fp16)[name = tensor<string, []>("linear_84_cast_fp16")];
            tensor<int32, [4]> var_2228 = const()[name = tensor<string, []>("op_2228"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_491_cast_fp16 = reshape(shape = var_2228, x = linear_84_cast_fp16)[name = tensor<string, []>("x_491_cast_fp16")];
            tensor<string, []> x_491_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_491_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2198_promoted_1 = const()[name = tensor<string, []>("op_2198_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_491_cast_fp16_to_fp32 = cast(dtype = x_491_cast_fp16_to_fp32_dtype_0, x = x_491_cast_fp16)[name = tensor<string, []>("cast_466")];
            tensor<fp32, [1, 1, 12, 128]> var_2232 = pow(x = x_491_cast_fp16_to_fp32, y = var_2198_promoted_1)[name = tensor<string, []>("op_2232")];
            tensor<int32, [1]> var_99_axes_0 = const()[name = tensor<string, []>("var_99_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_99_keep_dims_0 = const()[name = tensor<string, []>("var_99_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_99 = reduce_mean(axes = var_99_axes_0, keep_dims = var_99_keep_dims_0, x = var_2232)[name = tensor<string, []>("var_99")];
            tensor<string, []> var_99_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_99_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2236_to_fp16 = const()[name = tensor<string, []>("op_2236_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_99_to_fp16 = cast(dtype = var_99_to_fp16_dtype_0, x = var_99)[name = tensor<string, []>("cast_465")];
            tensor<fp16, [1, 1, 12, 1]> var_2237_cast_fp16 = add(x = var_99_to_fp16, y = var_2236_to_fp16)[name = tensor<string, []>("op_2237_cast_fp16")];
            tensor<string, []> var_2237_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2237_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2238_epsilon_0 = const()[name = tensor<string, []>("op_2238_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_2237_cast_fp16_to_fp32 = cast(dtype = var_2237_cast_fp16_to_fp32_dtype_0, x = var_2237_cast_fp16)[name = tensor<string, []>("cast_464")];
            tensor<fp32, [1, 1, 12, 1]> var_2238 = rsqrt(epsilon = var_2238_epsilon_0, x = var_2237_cast_fp16_to_fp32)[name = tensor<string, []>("op_2238")];
            tensor<string, []> var_2238_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2238_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_2238_to_fp16 = cast(dtype = var_2238_to_fp16_dtype_0, x = var_2238)[name = tensor<string, []>("cast_463")];
            tensor<fp16, [1, 1, 12, 128]> x_497_cast_fp16 = mul(x = x_491_cast_fp16, y = var_2238_to_fp16)[name = tensor<string, []>("x_497_cast_fp16")];
            tensor<fp16, [128]> layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331026368)))];
            tensor<fp16, [1, 1, 12, 128]> var_2240_cast_fp16 = mul(x = layers_12_self_attn_q_norm_weight_to_fp16, y = x_497_cast_fp16)[name = tensor<string, []>("op_2240_cast_fp16")];
            tensor<int32, [4]> q_49_perm_0 = const()[name = tensor<string, []>("q_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331026688)))];
            tensor<fp16, [1, 1, 512]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = x_489_cast_fp16)[name = tensor<string, []>("linear_85_cast_fp16")];
            tensor<int32, [4]> var_2244 = const()[name = tensor<string, []>("op_2244"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_499_cast_fp16 = reshape(shape = var_2244, x = linear_85_cast_fp16)[name = tensor<string, []>("x_499_cast_fp16")];
            tensor<string, []> x_499_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_499_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2198_promoted_2 = const()[name = tensor<string, []>("op_2198_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_499_cast_fp16_to_fp32 = cast(dtype = x_499_cast_fp16_to_fp32_dtype_0, x = x_499_cast_fp16)[name = tensor<string, []>("cast_462")];
            tensor<fp32, [1, 1, 4, 128]> var_2248 = pow(x = x_499_cast_fp16_to_fp32, y = var_2198_promoted_2)[name = tensor<string, []>("op_2248")];
            tensor<int32, [1]> var_101_axes_0 = const()[name = tensor<string, []>("var_101_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_101_keep_dims_0 = const()[name = tensor<string, []>("var_101_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_101 = reduce_mean(axes = var_101_axes_0, keep_dims = var_101_keep_dims_0, x = var_2248)[name = tensor<string, []>("var_101")];
            tensor<string, []> var_101_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_101_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2252_to_fp16 = const()[name = tensor<string, []>("op_2252_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_101_to_fp16 = cast(dtype = var_101_to_fp16_dtype_0, x = var_101)[name = tensor<string, []>("cast_461")];
            tensor<fp16, [1, 1, 4, 1]> var_2253_cast_fp16 = add(x = var_101_to_fp16, y = var_2252_to_fp16)[name = tensor<string, []>("op_2253_cast_fp16")];
            tensor<string, []> var_2253_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2253_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2254_epsilon_0 = const()[name = tensor<string, []>("op_2254_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_2253_cast_fp16_to_fp32 = cast(dtype = var_2253_cast_fp16_to_fp32_dtype_0, x = var_2253_cast_fp16)[name = tensor<string, []>("cast_460")];
            tensor<fp32, [1, 1, 4, 1]> var_2254 = rsqrt(epsilon = var_2254_epsilon_0, x = var_2253_cast_fp16_to_fp32)[name = tensor<string, []>("op_2254")];
            tensor<string, []> var_2254_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2254_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_2254_to_fp16 = cast(dtype = var_2254_to_fp16_dtype_0, x = var_2254)[name = tensor<string, []>("cast_459")];
            tensor<fp16, [1, 1, 4, 128]> x_505_cast_fp16 = mul(x = x_499_cast_fp16, y = var_2254_to_fp16)[name = tensor<string, []>("x_505_cast_fp16")];
            tensor<fp16, [128]> layers_12_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331551040)))];
            tensor<fp16, [1, 1, 4, 128]> var_2256_cast_fp16 = mul(x = layers_12_self_attn_k_norm_weight_to_fp16, y = x_505_cast_fp16)[name = tensor<string, []>("op_2256_cast_fp16")];
            tensor<int32, [4]> k_49_perm_0 = const()[name = tensor<string, []>("k_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331551360)))];
            tensor<fp16, [1, 1, 512]> linear_86_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = x_489_cast_fp16)[name = tensor<string, []>("linear_86_cast_fp16")];
            tensor<int32, [4]> var_2260 = const()[name = tensor<string, []>("op_2260"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_2261_cast_fp16 = reshape(shape = var_2260, x = linear_86_cast_fp16)[name = tensor<string, []>("op_2261_cast_fp16")];
            tensor<int32, [4]> v_25_perm_0 = const()[name = tensor<string, []>("v_25_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_49_cast_fp16 = transpose(perm = q_49_perm_0, x = var_2240_cast_fp16)[name = tensor<string, []>("transpose_63")];
            tensor<fp16, [1, 12, 1, 128]> var_2265_cast_fp16 = mul(x = q_49_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2265_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = tensor<string, []>("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = tensor<string, []>("x1_49_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = tensor<string, []>("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_49_cast_fp16)[name = tensor<string, []>("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = tensor<string, []>("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = tensor<string, []>("x2_49_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = tensor<string, []>("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_49_cast_fp16)[name = tensor<string, []>("x2_49_cast_fp16")];
            tensor<fp16, []> const_25_promoted_to_fp16 = const()[name = tensor<string, []>("const_25_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_2268_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_25_promoted_to_fp16)[name = tensor<string, []>("op_2268_cast_fp16")];
            tensor<bool, []> var_2270_interleave_0 = const()[name = tensor<string, []>("op_2270_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_2270_cast_fp16 = concat(axis = var_2199, interleave = var_2270_interleave_0, values = (var_2268_cast_fp16, x1_49_cast_fp16))[name = tensor<string, []>("op_2270_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_2271_cast_fp16 = mul(x = var_2270_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2271_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_51_cast_fp16 = add(x = var_2265_cast_fp16, y = var_2271_cast_fp16)[name = tensor<string, []>("q_51_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_49_cast_fp16 = transpose(perm = k_49_perm_0, x = var_2256_cast_fp16)[name = tensor<string, []>("transpose_62")];
            tensor<fp16, [1, 4, 1, 128]> var_2273_cast_fp16 = mul(x = k_49_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2273_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = tensor<string, []>("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = tensor<string, []>("x1_51_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = tensor<string, []>("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_49_cast_fp16)[name = tensor<string, []>("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = tensor<string, []>("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = tensor<string, []>("x2_51_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = tensor<string, []>("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_49_cast_fp16)[name = tensor<string, []>("x2_51_cast_fp16")];
            tensor<fp16, []> const_26_promoted_to_fp16 = const()[name = tensor<string, []>("const_26_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_2276_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_26_promoted_to_fp16)[name = tensor<string, []>("op_2276_cast_fp16")];
            tensor<bool, []> var_2278_interleave_0 = const()[name = tensor<string, []>("op_2278_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_2278_cast_fp16 = concat(axis = var_2199, interleave = var_2278_interleave_0, values = (var_2276_cast_fp16, x1_51_cast_fp16))[name = tensor<string, []>("op_2278_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_2279_cast_fp16 = mul(x = var_2278_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2279_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_51_cast_fp16 = add(x = var_2273_cast_fp16, y = var_2279_cast_fp16)[name = tensor<string, []>("k_51_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2282_cast_fp16 = mul(x = k_cache_25_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2282_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2283_cast_fp16 = mul(x = k_51_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2283_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_25_cast_fp16 = add(x = var_2282_cast_fp16, y = var_2283_cast_fp16)[name = tensor<string, []>("k_full_25_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2286_cast_fp16 = mul(x = v_cache_25_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2286_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_2261_cast_fp16)[name = tensor<string, []>("transpose_61")];
            tensor<fp16, [1, 4, 2048, 128]> var_2287_cast_fp16 = mul(x = v_25_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2287_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_25_cast_fp16 = add(x = var_2286_cast_fp16, y = var_2287_cast_fp16)[name = tensor<string, []>("v_full_25_cast_fp16")];
            tensor<int32, [1]> var_2289_axes_0 = const()[name = tensor<string, []>("op_2289_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2289_cast_fp16 = expand_dims(axes = var_2289_axes_0, x = k_full_25_cast_fp16)[name = tensor<string, []>("op_2289_cast_fp16")];
            tensor<int32, [5]> var_2291_reps_0 = const()[name = tensor<string, []>("op_2291_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2291_cast_fp16 = tile(reps = var_2291_reps_0, x = var_2289_cast_fp16)[name = tensor<string, []>("op_2291_cast_fp16")];
            tensor<int32, [4]> var_2292 = const()[name = tensor<string, []>("op_2292"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_25_cast_fp16 = reshape(shape = var_2292, x = var_2291_cast_fp16)[name = tensor<string, []>("k_rep_25_cast_fp16")];
            tensor<int32, [1]> var_2294_axes_0 = const()[name = tensor<string, []>("op_2294_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2294_cast_fp16 = expand_dims(axes = var_2294_axes_0, x = v_full_25_cast_fp16)[name = tensor<string, []>("op_2294_cast_fp16")];
            tensor<int32, [5]> var_2296_reps_0 = const()[name = tensor<string, []>("op_2296_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2296_cast_fp16 = tile(reps = var_2296_reps_0, x = var_2294_cast_fp16)[name = tensor<string, []>("op_2296_cast_fp16")];
            tensor<int32, [4]> var_2297 = const()[name = tensor<string, []>("op_2297"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_25_cast_fp16 = reshape(shape = var_2297, x = var_2296_cast_fp16)[name = tensor<string, []>("v_rep_25_cast_fp16")];
            tensor<bool, []> var_2300_transpose_x_1 = const()[name = tensor<string, []>("op_2300_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2300_transpose_y_1 = const()[name = tensor<string, []>("op_2300_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_2300_cast_fp16 = matmul(transpose_x = var_2300_transpose_x_1, transpose_y = var_2300_transpose_y_1, x = q_51_cast_fp16, y = k_rep_25_cast_fp16)[name = tensor<string, []>("op_2300_cast_fp16")];
            tensor<fp16, []> var_2301_to_fp16 = const()[name = tensor<string, []>("op_2301_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_49_cast_fp16 = mul(x = var_2300_cast_fp16, y = var_2301_to_fp16)[name = tensor<string, []>("attn_49_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_51_cast_fp16 = add(x = attn_49_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_51_cast_fp16")];
            tensor<string, []> input_51_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_51_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_51_cast_fp16_to_fp32 = cast(dtype = input_51_cast_fp16_to_fp32_dtype_0, x = input_51_cast_fp16)[name = tensor<string, []>("cast_458")];
            tensor<fp32, [1, 12, 1, 2048]> attn_51 = softmax(axis = var_2199, x = input_51_cast_fp16_to_fp32)[name = tensor<string, []>("attn_51")];
            tensor<bool, []> out_25_transpose_x_0 = const()[name = tensor<string, []>("out_25_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_25_transpose_y_0 = const()[name = tensor<string, []>("out_25_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_51_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_51_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_51_to_fp16 = cast(dtype = attn_51_to_fp16_dtype_0, x = attn_51)[name = tensor<string, []>("cast_457")];
            tensor<fp16, [1, 12, 1, 128]> out_25_cast_fp16 = matmul(transpose_x = out_25_transpose_x_0, transpose_y = out_25_transpose_y_0, x = attn_51_to_fp16, y = v_rep_25_cast_fp16)[name = tensor<string, []>("out_25_cast_fp16")];
            tensor<int32, [4]> var_2306_perm_0 = const()[name = tensor<string, []>("op_2306_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2308 = const()[name = tensor<string, []>("op_2308"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_2306_cast_fp16 = transpose(perm = var_2306_perm_0, x = out_25_cast_fp16)[name = tensor<string, []>("transpose_60")];
            tensor<fp16, [1, 1, 1536]> x_507_cast_fp16 = reshape(shape = var_2308, x = var_2306_cast_fp16)[name = tensor<string, []>("x_507_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332075712)))];
            tensor<fp16, [1, 1, 512]> linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = x_507_cast_fp16)[name = tensor<string, []>("linear_87_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_509_cast_fp16 = add(x = x_481_cast_fp16, y = linear_87_cast_fp16)[name = tensor<string, []>("x_509_cast_fp16")];
            tensor<string, []> x_509_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_509_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2198_promoted_3 = const()[name = tensor<string, []>("op_2198_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_509_cast_fp16_to_fp32 = cast(dtype = x_509_cast_fp16_to_fp32_dtype_0, x = x_509_cast_fp16)[name = tensor<string, []>("cast_456")];
            tensor<fp32, [1, 1, 512]> var_2319 = pow(x = x_509_cast_fp16_to_fp32, y = var_2198_promoted_3)[name = tensor<string, []>("op_2319")];
            tensor<int32, [1]> var_103_axes_0 = const()[name = tensor<string, []>("var_103_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_103_keep_dims_0 = const()[name = tensor<string, []>("var_103_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_103 = reduce_mean(axes = var_103_axes_0, keep_dims = var_103_keep_dims_0, x = var_2319)[name = tensor<string, []>("var_103")];
            tensor<string, []> var_103_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_103_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2323_to_fp16 = const()[name = tensor<string, []>("op_2323_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_103_to_fp16 = cast(dtype = var_103_to_fp16_dtype_0, x = var_103)[name = tensor<string, []>("cast_455")];
            tensor<fp16, [1, 1, 1]> var_2324_cast_fp16 = add(x = var_103_to_fp16, y = var_2323_to_fp16)[name = tensor<string, []>("op_2324_cast_fp16")];
            tensor<string, []> var_2324_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2324_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2325_epsilon_0 = const()[name = tensor<string, []>("op_2325_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2324_cast_fp16_to_fp32 = cast(dtype = var_2324_cast_fp16_to_fp32_dtype_0, x = var_2324_cast_fp16)[name = tensor<string, []>("cast_454")];
            tensor<fp32, [1, 1, 1]> var_2325 = rsqrt(epsilon = var_2325_epsilon_0, x = var_2324_cast_fp16_to_fp32)[name = tensor<string, []>("op_2325")];
            tensor<string, []> var_2325_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2325_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2325_to_fp16 = cast(dtype = var_2325_to_fp16_dtype_0, x = var_2325)[name = tensor<string, []>("cast_453")];
            tensor<fp16, [1, 1, 512]> x_515_cast_fp16 = mul(x = x_509_cast_fp16, y = var_2325_to_fp16)[name = tensor<string, []>("x_515_cast_fp16")];
            tensor<fp16, [512]> layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333648640)))];
            tensor<fp16, [1, 1, 512]> x_517_cast_fp16 = mul(x = layers_12_post_attention_layernorm_weight_to_fp16, y = x_515_cast_fp16)[name = tensor<string, []>("x_517_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_12_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(333649728)))];
            tensor<fp16, [1, 1, 1536]> linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_12_mlp_gate_proj_weight_to_fp16, x = x_517_cast_fp16)[name = tensor<string, []>("linear_88_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_2336_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor<string, []>("op_2336_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_12_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335222656)))];
            tensor<fp16, [1, 1, 1536]> linear_89_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_12_mlp_up_proj_weight_to_fp16, x = x_517_cast_fp16)[name = tensor<string, []>("linear_89_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_519_cast_fp16 = mul(x = var_2336_cast_fp16, y = linear_89_cast_fp16)[name = tensor<string, []>("x_519_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_12_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_12_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336795584)))];
            tensor<fp16, [1, 1, 512]> linear_90_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_12_mlp_down_proj_weight_to_fp16, x = x_519_cast_fp16)[name = tensor<string, []>("linear_90_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_521_cast_fp16 = add(x = x_509_cast_fp16, y = linear_90_cast_fp16)[name = tensor<string, []>("x_521_cast_fp16")];
            tensor<string, []> x_521_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_521_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_27_begin_0 = const()[name = tensor<string, []>("k_cache_27_begin_0"), val = tensor<int32, [5]>([13, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_27_end_0 = const()[name = tensor<string, []>("k_cache_27_end_0"), val = tensor<int32, [5]>([14, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_27_end_mask_0 = const()[name = tensor<string, []>("k_cache_27_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_27_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_27_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_27_cast_fp16 = slice_by_index(begin = k_cache_27_begin_0, end = k_cache_27_end_0, end_mask = k_cache_27_end_mask_0, squeeze_mask = k_cache_27_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_27_cast_fp16")];
            tensor<int32, [5]> v_cache_27_begin_0 = const()[name = tensor<string, []>("v_cache_27_begin_0"), val = tensor<int32, [5]>([13, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_27_end_0 = const()[name = tensor<string, []>("v_cache_27_end_0"), val = tensor<int32, [5]>([14, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_27_end_mask_0 = const()[name = tensor<string, []>("v_cache_27_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_27_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_27_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_27_cast_fp16 = slice_by_index(begin = v_cache_27_begin_0, end = v_cache_27_end_0, end_mask = v_cache_27_end_mask_0, squeeze_mask = v_cache_27_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_27_cast_fp16")];
            tensor<int32, []> var_2368 = const()[name = tensor<string, []>("op_2368"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_2367_promoted = const()[name = tensor<string, []>("op_2367_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_521_cast_fp16_to_fp32 = cast(dtype = x_521_cast_fp16_to_fp32_dtype_0, x = x_521_cast_fp16)[name = tensor<string, []>("cast_452")];
            tensor<fp32, [1, 1, 512]> var_2377 = pow(x = x_521_cast_fp16_to_fp32, y = var_2367_promoted)[name = tensor<string, []>("op_2377")];
            tensor<int32, [1]> var_105_axes_0 = const()[name = tensor<string, []>("var_105_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_105_keep_dims_0 = const()[name = tensor<string, []>("var_105_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_105_0 = reduce_mean(axes = var_105_axes_0, keep_dims = var_105_keep_dims_0, x = var_2377)[name = tensor<string, []>("var_105")];
            tensor<string, []> var_105_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_105_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2381_to_fp16 = const()[name = tensor<string, []>("op_2381_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_105_to_fp16 = cast(dtype = var_105_to_fp16_dtype_0, x = var_105_0)[name = tensor<string, []>("cast_451")];
            tensor<fp16, [1, 1, 1]> var_2382_cast_fp16 = add(x = var_105_to_fp16, y = var_2381_to_fp16)[name = tensor<string, []>("op_2382_cast_fp16")];
            tensor<string, []> var_2382_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2382_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2383_epsilon_0 = const()[name = tensor<string, []>("op_2383_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2382_cast_fp16_to_fp32 = cast(dtype = var_2382_cast_fp16_to_fp32_dtype_0, x = var_2382_cast_fp16)[name = tensor<string, []>("cast_450")];
            tensor<fp32, [1, 1, 1]> var_2383 = rsqrt(epsilon = var_2383_epsilon_0, x = var_2382_cast_fp16_to_fp32)[name = tensor<string, []>("op_2383")];
            tensor<string, []> var_2383_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2383_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2383_to_fp16 = cast(dtype = var_2383_to_fp16_dtype_0, x = var_2383)[name = tensor<string, []>("cast_449")];
            tensor<fp16, [1, 1, 512]> x_527_cast_fp16 = mul(x = x_521_cast_fp16, y = var_2383_to_fp16)[name = tensor<string, []>("x_527_cast_fp16")];
            tensor<fp16, [512]> layers_13_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338368512)))];
            tensor<fp16, [1, 1, 512]> x_529_cast_fp16 = mul(x = layers_13_input_layernorm_weight_to_fp16, y = x_527_cast_fp16)[name = tensor<string, []>("x_529_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(338369600)))];
            tensor<fp16, [1, 1, 1536]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = x_529_cast_fp16)[name = tensor<string, []>("linear_91_cast_fp16")];
            tensor<int32, [4]> var_2397 = const()[name = tensor<string, []>("op_2397"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_531_cast_fp16 = reshape(shape = var_2397, x = linear_91_cast_fp16)[name = tensor<string, []>("x_531_cast_fp16")];
            tensor<string, []> x_531_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_531_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2367_promoted_1 = const()[name = tensor<string, []>("op_2367_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_531_cast_fp16_to_fp32 = cast(dtype = x_531_cast_fp16_to_fp32_dtype_0, x = x_531_cast_fp16)[name = tensor<string, []>("cast_448")];
            tensor<fp32, [1, 1, 12, 128]> var_2401 = pow(x = x_531_cast_fp16_to_fp32, y = var_2367_promoted_1)[name = tensor<string, []>("op_2401")];
            tensor<int32, [1]> var_107_axes_0 = const()[name = tensor<string, []>("var_107_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_107_keep_dims_0 = const()[name = tensor<string, []>("var_107_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_107_0 = reduce_mean(axes = var_107_axes_0, keep_dims = var_107_keep_dims_0, x = var_2401)[name = tensor<string, []>("var_107")];
            tensor<string, []> var_107_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_107_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2405_to_fp16 = const()[name = tensor<string, []>("op_2405_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_107_to_fp16_0 = cast(dtype = var_107_to_fp16_dtype_0, x = var_107_0)[name = tensor<string, []>("cast_447")];
            tensor<fp16, [1, 1, 12, 1]> var_2406_cast_fp16 = add(x = var_107_to_fp16_0, y = var_2405_to_fp16)[name = tensor<string, []>("op_2406_cast_fp16")];
            tensor<string, []> var_2406_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2406_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2407_epsilon_0 = const()[name = tensor<string, []>("op_2407_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_2406_cast_fp16_to_fp32 = cast(dtype = var_2406_cast_fp16_to_fp32_dtype_0, x = var_2406_cast_fp16)[name = tensor<string, []>("cast_446")];
            tensor<fp32, [1, 1, 12, 1]> var_2407 = rsqrt(epsilon = var_2407_epsilon_0, x = var_2406_cast_fp16_to_fp32)[name = tensor<string, []>("op_2407")];
            tensor<string, []> var_2407_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2407_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_2407_to_fp16 = cast(dtype = var_2407_to_fp16_dtype_0, x = var_2407)[name = tensor<string, []>("cast_445")];
            tensor<fp16, [1, 1, 12, 128]> x_537_cast_fp16 = mul(x = x_531_cast_fp16, y = var_2407_to_fp16)[name = tensor<string, []>("x_537_cast_fp16")];
            tensor<fp16, [128]> layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(339942528)))];
            tensor<fp16, [1, 1, 12, 128]> var_2409_cast_fp16 = mul(x = layers_13_self_attn_q_norm_weight_to_fp16, y = x_537_cast_fp16)[name = tensor<string, []>("op_2409_cast_fp16")];
            tensor<int32, [4]> q_53_perm_0 = const()[name = tensor<string, []>("q_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(339942848)))];
            tensor<fp16, [1, 1, 512]> linear_92_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = x_529_cast_fp16)[name = tensor<string, []>("linear_92_cast_fp16")];
            tensor<int32, [4]> var_2413 = const()[name = tensor<string, []>("op_2413"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_539_cast_fp16 = reshape(shape = var_2413, x = linear_92_cast_fp16)[name = tensor<string, []>("x_539_cast_fp16")];
            tensor<string, []> x_539_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_539_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2367_promoted_2 = const()[name = tensor<string, []>("op_2367_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_539_cast_fp16_to_fp32 = cast(dtype = x_539_cast_fp16_to_fp32_dtype_0, x = x_539_cast_fp16)[name = tensor<string, []>("cast_444")];
            tensor<fp32, [1, 1, 4, 128]> var_2417 = pow(x = x_539_cast_fp16_to_fp32, y = var_2367_promoted_2)[name = tensor<string, []>("op_2417")];
            tensor<int32, [1]> var_109_axes_0 = const()[name = tensor<string, []>("var_109_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_109_keep_dims_0 = const()[name = tensor<string, []>("var_109_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_109 = reduce_mean(axes = var_109_axes_0, keep_dims = var_109_keep_dims_0, x = var_2417)[name = tensor<string, []>("var_109")];
            tensor<string, []> var_109_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_109_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2421_to_fp16 = const()[name = tensor<string, []>("op_2421_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_109_to_fp16 = cast(dtype = var_109_to_fp16_dtype_0, x = var_109)[name = tensor<string, []>("cast_443")];
            tensor<fp16, [1, 1, 4, 1]> var_2422_cast_fp16 = add(x = var_109_to_fp16, y = var_2421_to_fp16)[name = tensor<string, []>("op_2422_cast_fp16")];
            tensor<string, []> var_2422_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2422_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2423_epsilon_0 = const()[name = tensor<string, []>("op_2423_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_2422_cast_fp16_to_fp32 = cast(dtype = var_2422_cast_fp16_to_fp32_dtype_0, x = var_2422_cast_fp16)[name = tensor<string, []>("cast_442")];
            tensor<fp32, [1, 1, 4, 1]> var_2423 = rsqrt(epsilon = var_2423_epsilon_0, x = var_2422_cast_fp16_to_fp32)[name = tensor<string, []>("op_2423")];
            tensor<string, []> var_2423_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2423_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_2423_to_fp16 = cast(dtype = var_2423_to_fp16_dtype_0, x = var_2423)[name = tensor<string, []>("cast_441")];
            tensor<fp16, [1, 1, 4, 128]> x_545_cast_fp16 = mul(x = x_539_cast_fp16, y = var_2423_to_fp16)[name = tensor<string, []>("x_545_cast_fp16")];
            tensor<fp16, [128]> layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340467200)))];
            tensor<fp16, [1, 1, 4, 128]> var_2425_cast_fp16 = mul(x = layers_13_self_attn_k_norm_weight_to_fp16, y = x_545_cast_fp16)[name = tensor<string, []>("op_2425_cast_fp16")];
            tensor<int32, [4]> k_53_perm_0 = const()[name = tensor<string, []>("k_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340467520)))];
            tensor<fp16, [1, 1, 512]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = x_529_cast_fp16)[name = tensor<string, []>("linear_93_cast_fp16")];
            tensor<int32, [4]> var_2429 = const()[name = tensor<string, []>("op_2429"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_2430_cast_fp16 = reshape(shape = var_2429, x = linear_93_cast_fp16)[name = tensor<string, []>("op_2430_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = tensor<string, []>("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_53_cast_fp16 = transpose(perm = q_53_perm_0, x = var_2409_cast_fp16)[name = tensor<string, []>("transpose_59")];
            tensor<fp16, [1, 12, 1, 128]> var_2434_cast_fp16 = mul(x = q_53_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2434_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = tensor<string, []>("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = tensor<string, []>("x1_53_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = tensor<string, []>("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_53_cast_fp16)[name = tensor<string, []>("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = tensor<string, []>("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = tensor<string, []>("x2_53_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = tensor<string, []>("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_53_cast_fp16)[name = tensor<string, []>("x2_53_cast_fp16")];
            tensor<fp16, []> const_27_promoted_to_fp16 = const()[name = tensor<string, []>("const_27_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_2437_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_27_promoted_to_fp16)[name = tensor<string, []>("op_2437_cast_fp16")];
            tensor<bool, []> var_2439_interleave_0 = const()[name = tensor<string, []>("op_2439_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_2439_cast_fp16 = concat(axis = var_2368, interleave = var_2439_interleave_0, values = (var_2437_cast_fp16, x1_53_cast_fp16))[name = tensor<string, []>("op_2439_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_2440_cast_fp16 = mul(x = var_2439_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2440_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_55_cast_fp16 = add(x = var_2434_cast_fp16, y = var_2440_cast_fp16)[name = tensor<string, []>("q_55_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_53_cast_fp16 = transpose(perm = k_53_perm_0, x = var_2425_cast_fp16)[name = tensor<string, []>("transpose_58")];
            tensor<fp16, [1, 4, 1, 128]> var_2442_cast_fp16 = mul(x = k_53_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2442_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = tensor<string, []>("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = tensor<string, []>("x1_55_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = tensor<string, []>("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_53_cast_fp16)[name = tensor<string, []>("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = tensor<string, []>("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = tensor<string, []>("x2_55_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = tensor<string, []>("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_53_cast_fp16)[name = tensor<string, []>("x2_55_cast_fp16")];
            tensor<fp16, []> const_28_promoted_to_fp16 = const()[name = tensor<string, []>("const_28_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_2445_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_28_promoted_to_fp16)[name = tensor<string, []>("op_2445_cast_fp16")];
            tensor<bool, []> var_2447_interleave_0 = const()[name = tensor<string, []>("op_2447_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_2447_cast_fp16 = concat(axis = var_2368, interleave = var_2447_interleave_0, values = (var_2445_cast_fp16, x1_55_cast_fp16))[name = tensor<string, []>("op_2447_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_2448_cast_fp16 = mul(x = var_2447_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2448_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_55_cast_fp16 = add(x = var_2442_cast_fp16, y = var_2448_cast_fp16)[name = tensor<string, []>("k_55_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2451_cast_fp16 = mul(x = k_cache_27_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2451_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2452_cast_fp16 = mul(x = k_55_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2452_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_27_cast_fp16 = add(x = var_2451_cast_fp16, y = var_2452_cast_fp16)[name = tensor<string, []>("k_full_27_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2455_cast_fp16 = mul(x = v_cache_27_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2455_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_2430_cast_fp16)[name = tensor<string, []>("transpose_57")];
            tensor<fp16, [1, 4, 2048, 128]> var_2456_cast_fp16 = mul(x = v_27_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2456_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_27_cast_fp16 = add(x = var_2455_cast_fp16, y = var_2456_cast_fp16)[name = tensor<string, []>("v_full_27_cast_fp16")];
            tensor<int32, [1]> var_2458_axes_0 = const()[name = tensor<string, []>("op_2458_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2458_cast_fp16 = expand_dims(axes = var_2458_axes_0, x = k_full_27_cast_fp16)[name = tensor<string, []>("op_2458_cast_fp16")];
            tensor<int32, [5]> var_2460_reps_0 = const()[name = tensor<string, []>("op_2460_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2460_cast_fp16 = tile(reps = var_2460_reps_0, x = var_2458_cast_fp16)[name = tensor<string, []>("op_2460_cast_fp16")];
            tensor<int32, [4]> var_2461 = const()[name = tensor<string, []>("op_2461"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_27_cast_fp16 = reshape(shape = var_2461, x = var_2460_cast_fp16)[name = tensor<string, []>("k_rep_27_cast_fp16")];
            tensor<int32, [1]> var_2463_axes_0 = const()[name = tensor<string, []>("op_2463_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2463_cast_fp16 = expand_dims(axes = var_2463_axes_0, x = v_full_27_cast_fp16)[name = tensor<string, []>("op_2463_cast_fp16")];
            tensor<int32, [5]> var_2465_reps_0 = const()[name = tensor<string, []>("op_2465_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2465_cast_fp16 = tile(reps = var_2465_reps_0, x = var_2463_cast_fp16)[name = tensor<string, []>("op_2465_cast_fp16")];
            tensor<int32, [4]> var_2466 = const()[name = tensor<string, []>("op_2466"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_27_cast_fp16 = reshape(shape = var_2466, x = var_2465_cast_fp16)[name = tensor<string, []>("v_rep_27_cast_fp16")];
            tensor<bool, []> var_2469_transpose_x_1 = const()[name = tensor<string, []>("op_2469_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2469_transpose_y_1 = const()[name = tensor<string, []>("op_2469_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_2469_cast_fp16 = matmul(transpose_x = var_2469_transpose_x_1, transpose_y = var_2469_transpose_y_1, x = q_55_cast_fp16, y = k_rep_27_cast_fp16)[name = tensor<string, []>("op_2469_cast_fp16")];
            tensor<fp16, []> var_2470_to_fp16 = const()[name = tensor<string, []>("op_2470_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_53_cast_fp16 = mul(x = var_2469_cast_fp16, y = var_2470_to_fp16)[name = tensor<string, []>("attn_53_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_55_cast_fp16 = add(x = attn_53_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_55_cast_fp16")];
            tensor<string, []> input_55_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_55_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_55_cast_fp16_to_fp32 = cast(dtype = input_55_cast_fp16_to_fp32_dtype_0, x = input_55_cast_fp16)[name = tensor<string, []>("cast_440")];
            tensor<fp32, [1, 12, 1, 2048]> attn_55 = softmax(axis = var_2368, x = input_55_cast_fp16_to_fp32)[name = tensor<string, []>("attn_55")];
            tensor<bool, []> out_27_transpose_x_0 = const()[name = tensor<string, []>("out_27_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_27_transpose_y_0 = const()[name = tensor<string, []>("out_27_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_55_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_55_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_55_to_fp16 = cast(dtype = attn_55_to_fp16_dtype_0, x = attn_55)[name = tensor<string, []>("cast_439")];
            tensor<fp16, [1, 12, 1, 128]> out_27_cast_fp16 = matmul(transpose_x = out_27_transpose_x_0, transpose_y = out_27_transpose_y_0, x = attn_55_to_fp16, y = v_rep_27_cast_fp16)[name = tensor<string, []>("out_27_cast_fp16")];
            tensor<int32, [4]> var_2475_perm_0 = const()[name = tensor<string, []>("op_2475_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2477 = const()[name = tensor<string, []>("op_2477"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_2475_cast_fp16 = transpose(perm = var_2475_perm_0, x = out_27_cast_fp16)[name = tensor<string, []>("transpose_56")];
            tensor<fp16, [1, 1, 1536]> x_547_cast_fp16 = reshape(shape = var_2477, x = var_2475_cast_fp16)[name = tensor<string, []>("x_547_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340991872)))];
            tensor<fp16, [1, 1, 512]> linear_94_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = x_547_cast_fp16)[name = tensor<string, []>("linear_94_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_549_cast_fp16 = add(x = x_521_cast_fp16, y = linear_94_cast_fp16)[name = tensor<string, []>("x_549_cast_fp16")];
            tensor<string, []> x_549_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_549_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2367_promoted_3 = const()[name = tensor<string, []>("op_2367_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_549_cast_fp16_to_fp32 = cast(dtype = x_549_cast_fp16_to_fp32_dtype_0, x = x_549_cast_fp16)[name = tensor<string, []>("cast_438")];
            tensor<fp32, [1, 1, 512]> var_2488 = pow(x = x_549_cast_fp16_to_fp32, y = var_2367_promoted_3)[name = tensor<string, []>("op_2488")];
            tensor<int32, [1]> var_111_axes_0 = const()[name = tensor<string, []>("var_111_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_111_keep_dims_0 = const()[name = tensor<string, []>("var_111_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_111 = reduce_mean(axes = var_111_axes_0, keep_dims = var_111_keep_dims_0, x = var_2488)[name = tensor<string, []>("var_111")];
            tensor<string, []> var_111_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_111_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2492_to_fp16 = const()[name = tensor<string, []>("op_2492_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_111_to_fp16 = cast(dtype = var_111_to_fp16_dtype_0, x = var_111)[name = tensor<string, []>("cast_437")];
            tensor<fp16, [1, 1, 1]> var_2493_cast_fp16 = add(x = var_111_to_fp16, y = var_2492_to_fp16)[name = tensor<string, []>("op_2493_cast_fp16")];
            tensor<string, []> var_2493_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2493_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2494_epsilon_0 = const()[name = tensor<string, []>("op_2494_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2493_cast_fp16_to_fp32 = cast(dtype = var_2493_cast_fp16_to_fp32_dtype_0, x = var_2493_cast_fp16)[name = tensor<string, []>("cast_436")];
            tensor<fp32, [1, 1, 1]> var_2494 = rsqrt(epsilon = var_2494_epsilon_0, x = var_2493_cast_fp16_to_fp32)[name = tensor<string, []>("op_2494")];
            tensor<string, []> var_2494_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2494_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2494_to_fp16 = cast(dtype = var_2494_to_fp16_dtype_0, x = var_2494)[name = tensor<string, []>("cast_435")];
            tensor<fp16, [1, 1, 512]> x_555_cast_fp16 = mul(x = x_549_cast_fp16, y = var_2494_to_fp16)[name = tensor<string, []>("x_555_cast_fp16")];
            tensor<fp16, [512]> layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342564800)))];
            tensor<fp16, [1, 1, 512]> x_557_cast_fp16 = mul(x = layers_13_post_attention_layernorm_weight_to_fp16, y = x_555_cast_fp16)[name = tensor<string, []>("x_557_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_13_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(342565888)))];
            tensor<fp16, [1, 1, 1536]> linear_95_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_13_mlp_gate_proj_weight_to_fp16, x = x_557_cast_fp16)[name = tensor<string, []>("linear_95_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_2505_cast_fp16 = silu(x = linear_95_cast_fp16)[name = tensor<string, []>("op_2505_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(344138816)))];
            tensor<fp16, [1, 1, 1536]> linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_13_mlp_up_proj_weight_to_fp16, x = x_557_cast_fp16)[name = tensor<string, []>("linear_96_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_559_cast_fp16 = mul(x = var_2505_cast_fp16, y = linear_96_cast_fp16)[name = tensor<string, []>("x_559_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_13_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_13_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(345711744)))];
            tensor<fp16, [1, 1, 512]> linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_13_mlp_down_proj_weight_to_fp16, x = x_559_cast_fp16)[name = tensor<string, []>("linear_97_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_561_cast_fp16 = add(x = x_549_cast_fp16, y = linear_97_cast_fp16)[name = tensor<string, []>("x_561_cast_fp16")];
            tensor<string, []> x_561_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_561_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_29_begin_0 = const()[name = tensor<string, []>("k_cache_29_begin_0"), val = tensor<int32, [5]>([14, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_29_end_0 = const()[name = tensor<string, []>("k_cache_29_end_0"), val = tensor<int32, [5]>([15, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_29_end_mask_0 = const()[name = tensor<string, []>("k_cache_29_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_29_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_29_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_29_cast_fp16 = slice_by_index(begin = k_cache_29_begin_0, end = k_cache_29_end_0, end_mask = k_cache_29_end_mask_0, squeeze_mask = k_cache_29_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_29_cast_fp16")];
            tensor<int32, [5]> v_cache_29_begin_0 = const()[name = tensor<string, []>("v_cache_29_begin_0"), val = tensor<int32, [5]>([14, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_29_end_0 = const()[name = tensor<string, []>("v_cache_29_end_0"), val = tensor<int32, [5]>([15, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_29_end_mask_0 = const()[name = tensor<string, []>("v_cache_29_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_29_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_29_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_29_cast_fp16 = slice_by_index(begin = v_cache_29_begin_0, end = v_cache_29_end_0, end_mask = v_cache_29_end_mask_0, squeeze_mask = v_cache_29_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_29_cast_fp16")];
            tensor<int32, []> var_2537 = const()[name = tensor<string, []>("op_2537"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_2536_promoted = const()[name = tensor<string, []>("op_2536_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_561_cast_fp16_to_fp32 = cast(dtype = x_561_cast_fp16_to_fp32_dtype_0, x = x_561_cast_fp16)[name = tensor<string, []>("cast_434")];
            tensor<fp32, [1, 1, 512]> var_2546 = pow(x = x_561_cast_fp16_to_fp32, y = var_2536_promoted)[name = tensor<string, []>("op_2546")];
            tensor<int32, [1]> var_113_axes_0 = const()[name = tensor<string, []>("var_113_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_113_keep_dims_0 = const()[name = tensor<string, []>("var_113_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_113 = reduce_mean(axes = var_113_axes_0, keep_dims = var_113_keep_dims_0, x = var_2546)[name = tensor<string, []>("var_113")];
            tensor<string, []> var_113_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_113_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2550_to_fp16 = const()[name = tensor<string, []>("op_2550_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_113_to_fp16 = cast(dtype = var_113_to_fp16_dtype_0, x = var_113)[name = tensor<string, []>("cast_433")];
            tensor<fp16, [1, 1, 1]> var_2551_cast_fp16 = add(x = var_113_to_fp16, y = var_2550_to_fp16)[name = tensor<string, []>("op_2551_cast_fp16")];
            tensor<string, []> var_2551_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2551_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2552_epsilon_0 = const()[name = tensor<string, []>("op_2552_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2551_cast_fp16_to_fp32 = cast(dtype = var_2551_cast_fp16_to_fp32_dtype_0, x = var_2551_cast_fp16)[name = tensor<string, []>("cast_432")];
            tensor<fp32, [1, 1, 1]> var_2552 = rsqrt(epsilon = var_2552_epsilon_0, x = var_2551_cast_fp16_to_fp32)[name = tensor<string, []>("op_2552")];
            tensor<string, []> var_2552_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2552_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2552_to_fp16 = cast(dtype = var_2552_to_fp16_dtype_0, x = var_2552)[name = tensor<string, []>("cast_431")];
            tensor<fp16, [1, 1, 512]> x_567_cast_fp16 = mul(x = x_561_cast_fp16, y = var_2552_to_fp16)[name = tensor<string, []>("x_567_cast_fp16")];
            tensor<fp16, [512]> layers_14_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347284672)))];
            tensor<fp16, [1, 1, 512]> x_569_cast_fp16 = mul(x = layers_14_input_layernorm_weight_to_fp16, y = x_567_cast_fp16)[name = tensor<string, []>("x_569_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(347285760)))];
            tensor<fp16, [1, 1, 1536]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = x_569_cast_fp16)[name = tensor<string, []>("linear_98_cast_fp16")];
            tensor<int32, [4]> var_2566 = const()[name = tensor<string, []>("op_2566"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_571_cast_fp16 = reshape(shape = var_2566, x = linear_98_cast_fp16)[name = tensor<string, []>("x_571_cast_fp16")];
            tensor<string, []> x_571_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_571_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2536_promoted_1 = const()[name = tensor<string, []>("op_2536_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_571_cast_fp16_to_fp32 = cast(dtype = x_571_cast_fp16_to_fp32_dtype_0, x = x_571_cast_fp16)[name = tensor<string, []>("cast_430")];
            tensor<fp32, [1, 1, 12, 128]> var_2570 = pow(x = x_571_cast_fp16_to_fp32, y = var_2536_promoted_1)[name = tensor<string, []>("op_2570")];
            tensor<int32, [1]> var_115_axes_0 = const()[name = tensor<string, []>("var_115_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_115_keep_dims_0 = const()[name = tensor<string, []>("var_115_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_115 = reduce_mean(axes = var_115_axes_0, keep_dims = var_115_keep_dims_0, x = var_2570)[name = tensor<string, []>("var_115")];
            tensor<string, []> var_115_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_115_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2574_to_fp16 = const()[name = tensor<string, []>("op_2574_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_115_to_fp16 = cast(dtype = var_115_to_fp16_dtype_0, x = var_115)[name = tensor<string, []>("cast_429")];
            tensor<fp16, [1, 1, 12, 1]> var_2575_cast_fp16 = add(x = var_115_to_fp16, y = var_2574_to_fp16)[name = tensor<string, []>("op_2575_cast_fp16")];
            tensor<string, []> var_2575_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2575_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2576_epsilon_0 = const()[name = tensor<string, []>("op_2576_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_2575_cast_fp16_to_fp32 = cast(dtype = var_2575_cast_fp16_to_fp32_dtype_0, x = var_2575_cast_fp16)[name = tensor<string, []>("cast_428")];
            tensor<fp32, [1, 1, 12, 1]> var_2576 = rsqrt(epsilon = var_2576_epsilon_0, x = var_2575_cast_fp16_to_fp32)[name = tensor<string, []>("op_2576")];
            tensor<string, []> var_2576_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2576_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_2576_to_fp16 = cast(dtype = var_2576_to_fp16_dtype_0, x = var_2576)[name = tensor<string, []>("cast_427")];
            tensor<fp16, [1, 1, 12, 128]> x_577_cast_fp16 = mul(x = x_571_cast_fp16, y = var_2576_to_fp16)[name = tensor<string, []>("x_577_cast_fp16")];
            tensor<fp16, [128]> layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348858688)))];
            tensor<fp16, [1, 1, 12, 128]> var_2578_cast_fp16 = mul(x = layers_14_self_attn_q_norm_weight_to_fp16, y = x_577_cast_fp16)[name = tensor<string, []>("op_2578_cast_fp16")];
            tensor<int32, [4]> q_57_perm_0 = const()[name = tensor<string, []>("q_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348859008)))];
            tensor<fp16, [1, 1, 512]> linear_99_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = x_569_cast_fp16)[name = tensor<string, []>("linear_99_cast_fp16")];
            tensor<int32, [4]> var_2582 = const()[name = tensor<string, []>("op_2582"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_579_cast_fp16 = reshape(shape = var_2582, x = linear_99_cast_fp16)[name = tensor<string, []>("x_579_cast_fp16")];
            tensor<string, []> x_579_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_579_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2536_promoted_2 = const()[name = tensor<string, []>("op_2536_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_579_cast_fp16_to_fp32 = cast(dtype = x_579_cast_fp16_to_fp32_dtype_0, x = x_579_cast_fp16)[name = tensor<string, []>("cast_426")];
            tensor<fp32, [1, 1, 4, 128]> var_2586 = pow(x = x_579_cast_fp16_to_fp32, y = var_2536_promoted_2)[name = tensor<string, []>("op_2586")];
            tensor<int32, [1]> var_117_axes_0 = const()[name = tensor<string, []>("var_117_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_117_keep_dims_0 = const()[name = tensor<string, []>("var_117_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_117 = reduce_mean(axes = var_117_axes_0, keep_dims = var_117_keep_dims_0, x = var_2586)[name = tensor<string, []>("var_117")];
            tensor<string, []> var_117_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_117_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2590_to_fp16 = const()[name = tensor<string, []>("op_2590_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_117_to_fp16 = cast(dtype = var_117_to_fp16_dtype_0, x = var_117)[name = tensor<string, []>("cast_425")];
            tensor<fp16, [1, 1, 4, 1]> var_2591_cast_fp16 = add(x = var_117_to_fp16, y = var_2590_to_fp16)[name = tensor<string, []>("op_2591_cast_fp16")];
            tensor<string, []> var_2591_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2591_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2592_epsilon_0 = const()[name = tensor<string, []>("op_2592_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_2591_cast_fp16_to_fp32 = cast(dtype = var_2591_cast_fp16_to_fp32_dtype_0, x = var_2591_cast_fp16)[name = tensor<string, []>("cast_424")];
            tensor<fp32, [1, 1, 4, 1]> var_2592 = rsqrt(epsilon = var_2592_epsilon_0, x = var_2591_cast_fp16_to_fp32)[name = tensor<string, []>("op_2592")];
            tensor<string, []> var_2592_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2592_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_2592_to_fp16 = cast(dtype = var_2592_to_fp16_dtype_0, x = var_2592)[name = tensor<string, []>("cast_423")];
            tensor<fp16, [1, 1, 4, 128]> x_585_cast_fp16 = mul(x = x_579_cast_fp16, y = var_2592_to_fp16)[name = tensor<string, []>("x_585_cast_fp16")];
            tensor<fp16, [128]> layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349383360)))];
            tensor<fp16, [1, 1, 4, 128]> var_2594_cast_fp16 = mul(x = layers_14_self_attn_k_norm_weight_to_fp16, y = x_585_cast_fp16)[name = tensor<string, []>("op_2594_cast_fp16")];
            tensor<int32, [4]> k_57_perm_0 = const()[name = tensor<string, []>("k_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349383680)))];
            tensor<fp16, [1, 1, 512]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = x_569_cast_fp16)[name = tensor<string, []>("linear_100_cast_fp16")];
            tensor<int32, [4]> var_2598 = const()[name = tensor<string, []>("op_2598"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_2599_cast_fp16 = reshape(shape = var_2598, x = linear_100_cast_fp16)[name = tensor<string, []>("op_2599_cast_fp16")];
            tensor<int32, [4]> v_29_perm_0 = const()[name = tensor<string, []>("v_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_57_cast_fp16 = transpose(perm = q_57_perm_0, x = var_2578_cast_fp16)[name = tensor<string, []>("transpose_55")];
            tensor<fp16, [1, 12, 1, 128]> var_2603_cast_fp16 = mul(x = q_57_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2603_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = tensor<string, []>("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = tensor<string, []>("x1_57_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = tensor<string, []>("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_57_cast_fp16)[name = tensor<string, []>("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = tensor<string, []>("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = tensor<string, []>("x2_57_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = tensor<string, []>("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_57_cast_fp16)[name = tensor<string, []>("x2_57_cast_fp16")];
            tensor<fp16, []> const_29_promoted_to_fp16 = const()[name = tensor<string, []>("const_29_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_2606_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_29_promoted_to_fp16)[name = tensor<string, []>("op_2606_cast_fp16")];
            tensor<bool, []> var_2608_interleave_0 = const()[name = tensor<string, []>("op_2608_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_2608_cast_fp16 = concat(axis = var_2537, interleave = var_2608_interleave_0, values = (var_2606_cast_fp16, x1_57_cast_fp16))[name = tensor<string, []>("op_2608_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_2609_cast_fp16 = mul(x = var_2608_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2609_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_59_cast_fp16 = add(x = var_2603_cast_fp16, y = var_2609_cast_fp16)[name = tensor<string, []>("q_59_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_57_cast_fp16 = transpose(perm = k_57_perm_0, x = var_2594_cast_fp16)[name = tensor<string, []>("transpose_54")];
            tensor<fp16, [1, 4, 1, 128]> var_2611_cast_fp16 = mul(x = k_57_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2611_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = tensor<string, []>("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = tensor<string, []>("x1_59_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = tensor<string, []>("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_57_cast_fp16)[name = tensor<string, []>("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = tensor<string, []>("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = tensor<string, []>("x2_59_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = tensor<string, []>("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_57_cast_fp16)[name = tensor<string, []>("x2_59_cast_fp16")];
            tensor<fp16, []> const_30_promoted_to_fp16 = const()[name = tensor<string, []>("const_30_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_2614_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_30_promoted_to_fp16)[name = tensor<string, []>("op_2614_cast_fp16")];
            tensor<bool, []> var_2616_interleave_0 = const()[name = tensor<string, []>("op_2616_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_2616_cast_fp16 = concat(axis = var_2537, interleave = var_2616_interleave_0, values = (var_2614_cast_fp16, x1_59_cast_fp16))[name = tensor<string, []>("op_2616_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_2617_cast_fp16 = mul(x = var_2616_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2617_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_59_cast_fp16 = add(x = var_2611_cast_fp16, y = var_2617_cast_fp16)[name = tensor<string, []>("k_59_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2620_cast_fp16 = mul(x = k_cache_29_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2620_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2621_cast_fp16 = mul(x = k_59_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2621_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_29_cast_fp16 = add(x = var_2620_cast_fp16, y = var_2621_cast_fp16)[name = tensor<string, []>("k_full_29_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2624_cast_fp16 = mul(x = v_cache_29_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2624_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_2599_cast_fp16)[name = tensor<string, []>("transpose_53")];
            tensor<fp16, [1, 4, 2048, 128]> var_2625_cast_fp16 = mul(x = v_29_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2625_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_29_cast_fp16 = add(x = var_2624_cast_fp16, y = var_2625_cast_fp16)[name = tensor<string, []>("v_full_29_cast_fp16")];
            tensor<int32, [1]> var_2627_axes_0 = const()[name = tensor<string, []>("op_2627_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2627_cast_fp16 = expand_dims(axes = var_2627_axes_0, x = k_full_29_cast_fp16)[name = tensor<string, []>("op_2627_cast_fp16")];
            tensor<int32, [5]> var_2629_reps_0 = const()[name = tensor<string, []>("op_2629_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2629_cast_fp16 = tile(reps = var_2629_reps_0, x = var_2627_cast_fp16)[name = tensor<string, []>("op_2629_cast_fp16")];
            tensor<int32, [4]> var_2630 = const()[name = tensor<string, []>("op_2630"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_29_cast_fp16 = reshape(shape = var_2630, x = var_2629_cast_fp16)[name = tensor<string, []>("k_rep_29_cast_fp16")];
            tensor<int32, [1]> var_2632_axes_0 = const()[name = tensor<string, []>("op_2632_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2632_cast_fp16 = expand_dims(axes = var_2632_axes_0, x = v_full_29_cast_fp16)[name = tensor<string, []>("op_2632_cast_fp16")];
            tensor<int32, [5]> var_2634_reps_0 = const()[name = tensor<string, []>("op_2634_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2634_cast_fp16 = tile(reps = var_2634_reps_0, x = var_2632_cast_fp16)[name = tensor<string, []>("op_2634_cast_fp16")];
            tensor<int32, [4]> var_2635 = const()[name = tensor<string, []>("op_2635"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_29_cast_fp16 = reshape(shape = var_2635, x = var_2634_cast_fp16)[name = tensor<string, []>("v_rep_29_cast_fp16")];
            tensor<bool, []> var_2638_transpose_x_1 = const()[name = tensor<string, []>("op_2638_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2638_transpose_y_1 = const()[name = tensor<string, []>("op_2638_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_2638_cast_fp16 = matmul(transpose_x = var_2638_transpose_x_1, transpose_y = var_2638_transpose_y_1, x = q_59_cast_fp16, y = k_rep_29_cast_fp16)[name = tensor<string, []>("op_2638_cast_fp16")];
            tensor<fp16, []> var_2639_to_fp16 = const()[name = tensor<string, []>("op_2639_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_57_cast_fp16 = mul(x = var_2638_cast_fp16, y = var_2639_to_fp16)[name = tensor<string, []>("attn_57_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_59_cast_fp16 = add(x = attn_57_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_59_cast_fp16")];
            tensor<string, []> input_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_59_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_59_cast_fp16_to_fp32 = cast(dtype = input_59_cast_fp16_to_fp32_dtype_0, x = input_59_cast_fp16)[name = tensor<string, []>("cast_422")];
            tensor<fp32, [1, 12, 1, 2048]> attn_59 = softmax(axis = var_2537, x = input_59_cast_fp16_to_fp32)[name = tensor<string, []>("attn_59")];
            tensor<bool, []> out_29_transpose_x_0 = const()[name = tensor<string, []>("out_29_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_29_transpose_y_0 = const()[name = tensor<string, []>("out_29_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_59_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_59_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_59_to_fp16 = cast(dtype = attn_59_to_fp16_dtype_0, x = attn_59)[name = tensor<string, []>("cast_421")];
            tensor<fp16, [1, 12, 1, 128]> out_29_cast_fp16 = matmul(transpose_x = out_29_transpose_x_0, transpose_y = out_29_transpose_y_0, x = attn_59_to_fp16, y = v_rep_29_cast_fp16)[name = tensor<string, []>("out_29_cast_fp16")];
            tensor<int32, [4]> var_2644_perm_0 = const()[name = tensor<string, []>("op_2644_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2646 = const()[name = tensor<string, []>("op_2646"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_2644_cast_fp16 = transpose(perm = var_2644_perm_0, x = out_29_cast_fp16)[name = tensor<string, []>("transpose_52")];
            tensor<fp16, [1, 1, 1536]> x_587_cast_fp16 = reshape(shape = var_2646, x = var_2644_cast_fp16)[name = tensor<string, []>("x_587_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(349908032)))];
            tensor<fp16, [1, 1, 512]> linear_101_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = x_587_cast_fp16)[name = tensor<string, []>("linear_101_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_589_cast_fp16 = add(x = x_561_cast_fp16, y = linear_101_cast_fp16)[name = tensor<string, []>("x_589_cast_fp16")];
            tensor<string, []> x_589_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_589_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2536_promoted_3 = const()[name = tensor<string, []>("op_2536_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_589_cast_fp16_to_fp32 = cast(dtype = x_589_cast_fp16_to_fp32_dtype_0, x = x_589_cast_fp16)[name = tensor<string, []>("cast_420")];
            tensor<fp32, [1, 1, 512]> var_2657 = pow(x = x_589_cast_fp16_to_fp32, y = var_2536_promoted_3)[name = tensor<string, []>("op_2657")];
            tensor<int32, [1]> var_119_axes_0 = const()[name = tensor<string, []>("var_119_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_119_keep_dims_0 = const()[name = tensor<string, []>("var_119_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_119 = reduce_mean(axes = var_119_axes_0, keep_dims = var_119_keep_dims_0, x = var_2657)[name = tensor<string, []>("var_119")];
            tensor<string, []> var_119_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_119_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2661_to_fp16 = const()[name = tensor<string, []>("op_2661_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_119_to_fp16 = cast(dtype = var_119_to_fp16_dtype_0, x = var_119)[name = tensor<string, []>("cast_419")];
            tensor<fp16, [1, 1, 1]> var_2662_cast_fp16 = add(x = var_119_to_fp16, y = var_2661_to_fp16)[name = tensor<string, []>("op_2662_cast_fp16")];
            tensor<string, []> var_2662_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2662_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2663_epsilon_0 = const()[name = tensor<string, []>("op_2663_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2662_cast_fp16_to_fp32 = cast(dtype = var_2662_cast_fp16_to_fp32_dtype_0, x = var_2662_cast_fp16)[name = tensor<string, []>("cast_418")];
            tensor<fp32, [1, 1, 1]> var_2663 = rsqrt(epsilon = var_2663_epsilon_0, x = var_2662_cast_fp16_to_fp32)[name = tensor<string, []>("op_2663")];
            tensor<string, []> var_2663_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2663_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2663_to_fp16 = cast(dtype = var_2663_to_fp16_dtype_0, x = var_2663)[name = tensor<string, []>("cast_417")];
            tensor<fp16, [1, 1, 512]> x_595_cast_fp16 = mul(x = x_589_cast_fp16, y = var_2663_to_fp16)[name = tensor<string, []>("x_595_cast_fp16")];
            tensor<fp16, [512]> layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351480960)))];
            tensor<fp16, [1, 1, 512]> x_597_cast_fp16 = mul(x = layers_14_post_attention_layernorm_weight_to_fp16, y = x_595_cast_fp16)[name = tensor<string, []>("x_597_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_14_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(351482048)))];
            tensor<fp16, [1, 1, 1536]> linear_102_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_14_mlp_gate_proj_weight_to_fp16, x = x_597_cast_fp16)[name = tensor<string, []>("linear_102_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_2674_cast_fp16 = silu(x = linear_102_cast_fp16)[name = tensor<string, []>("op_2674_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_14_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(353054976)))];
            tensor<fp16, [1, 1, 1536]> linear_103_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_14_mlp_up_proj_weight_to_fp16, x = x_597_cast_fp16)[name = tensor<string, []>("linear_103_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_599_cast_fp16 = mul(x = var_2674_cast_fp16, y = linear_103_cast_fp16)[name = tensor<string, []>("x_599_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_14_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_14_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(354627904)))];
            tensor<fp16, [1, 1, 512]> linear_104_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_14_mlp_down_proj_weight_to_fp16, x = x_599_cast_fp16)[name = tensor<string, []>("linear_104_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_601_cast_fp16 = add(x = x_589_cast_fp16, y = linear_104_cast_fp16)[name = tensor<string, []>("x_601_cast_fp16")];
            tensor<string, []> x_601_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_601_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_31_begin_0 = const()[name = tensor<string, []>("k_cache_31_begin_0"), val = tensor<int32, [5]>([15, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_31_end_0 = const()[name = tensor<string, []>("k_cache_31_end_0"), val = tensor<int32, [5]>([16, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_31_end_mask_0 = const()[name = tensor<string, []>("k_cache_31_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_31_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_31_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_31_cast_fp16 = slice_by_index(begin = k_cache_31_begin_0, end = k_cache_31_end_0, end_mask = k_cache_31_end_mask_0, squeeze_mask = k_cache_31_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_31_cast_fp16")];
            tensor<int32, [5]> v_cache_31_begin_0 = const()[name = tensor<string, []>("v_cache_31_begin_0"), val = tensor<int32, [5]>([15, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_31_end_0 = const()[name = tensor<string, []>("v_cache_31_end_0"), val = tensor<int32, [5]>([16, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_31_end_mask_0 = const()[name = tensor<string, []>("v_cache_31_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_31_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_31_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_31_cast_fp16 = slice_by_index(begin = v_cache_31_begin_0, end = v_cache_31_end_0, end_mask = v_cache_31_end_mask_0, squeeze_mask = v_cache_31_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_31_cast_fp16")];
            tensor<int32, []> var_2706 = const()[name = tensor<string, []>("op_2706"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_2705_promoted = const()[name = tensor<string, []>("op_2705_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_601_cast_fp16_to_fp32 = cast(dtype = x_601_cast_fp16_to_fp32_dtype_0, x = x_601_cast_fp16)[name = tensor<string, []>("cast_416")];
            tensor<fp32, [1, 1, 512]> var_2715 = pow(x = x_601_cast_fp16_to_fp32, y = var_2705_promoted)[name = tensor<string, []>("op_2715")];
            tensor<int32, [1]> var_121_axes_0 = const()[name = tensor<string, []>("var_121_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_121_keep_dims_0 = const()[name = tensor<string, []>("var_121_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_121 = reduce_mean(axes = var_121_axes_0, keep_dims = var_121_keep_dims_0, x = var_2715)[name = tensor<string, []>("var_121")];
            tensor<string, []> var_121_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_121_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2719_to_fp16 = const()[name = tensor<string, []>("op_2719_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_121_to_fp16 = cast(dtype = var_121_to_fp16_dtype_0, x = var_121)[name = tensor<string, []>("cast_415")];
            tensor<fp16, [1, 1, 1]> var_2720_cast_fp16 = add(x = var_121_to_fp16, y = var_2719_to_fp16)[name = tensor<string, []>("op_2720_cast_fp16")];
            tensor<string, []> var_2720_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2720_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2721_epsilon_0 = const()[name = tensor<string, []>("op_2721_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2720_cast_fp16_to_fp32 = cast(dtype = var_2720_cast_fp16_to_fp32_dtype_0, x = var_2720_cast_fp16)[name = tensor<string, []>("cast_414")];
            tensor<fp32, [1, 1, 1]> var_2721 = rsqrt(epsilon = var_2721_epsilon_0, x = var_2720_cast_fp16_to_fp32)[name = tensor<string, []>("op_2721")];
            tensor<string, []> var_2721_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2721_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2721_to_fp16 = cast(dtype = var_2721_to_fp16_dtype_0, x = var_2721)[name = tensor<string, []>("cast_413")];
            tensor<fp16, [1, 1, 512]> x_607_cast_fp16 = mul(x = x_601_cast_fp16, y = var_2721_to_fp16)[name = tensor<string, []>("x_607_cast_fp16")];
            tensor<fp16, [512]> layers_15_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356200832)))];
            tensor<fp16, [1, 1, 512]> x_609_cast_fp16 = mul(x = layers_15_input_layernorm_weight_to_fp16, y = x_607_cast_fp16)[name = tensor<string, []>("x_609_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356201920)))];
            tensor<fp16, [1, 1, 1536]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = x_609_cast_fp16)[name = tensor<string, []>("linear_105_cast_fp16")];
            tensor<int32, [4]> var_2735 = const()[name = tensor<string, []>("op_2735"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_611_cast_fp16 = reshape(shape = var_2735, x = linear_105_cast_fp16)[name = tensor<string, []>("x_611_cast_fp16")];
            tensor<string, []> x_611_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_611_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2705_promoted_1 = const()[name = tensor<string, []>("op_2705_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_611_cast_fp16_to_fp32 = cast(dtype = x_611_cast_fp16_to_fp32_dtype_0, x = x_611_cast_fp16)[name = tensor<string, []>("cast_412")];
            tensor<fp32, [1, 1, 12, 128]> var_2739 = pow(x = x_611_cast_fp16_to_fp32, y = var_2705_promoted_1)[name = tensor<string, []>("op_2739")];
            tensor<int32, [1]> var_123_axes_0 = const()[name = tensor<string, []>("var_123_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_123_keep_dims_0 = const()[name = tensor<string, []>("var_123_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_123 = reduce_mean(axes = var_123_axes_0, keep_dims = var_123_keep_dims_0, x = var_2739)[name = tensor<string, []>("var_123")];
            tensor<string, []> var_123_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_123_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2743_to_fp16 = const()[name = tensor<string, []>("op_2743_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_123_to_fp16 = cast(dtype = var_123_to_fp16_dtype_0, x = var_123)[name = tensor<string, []>("cast_411")];
            tensor<fp16, [1, 1, 12, 1]> var_2744_cast_fp16 = add(x = var_123_to_fp16, y = var_2743_to_fp16)[name = tensor<string, []>("op_2744_cast_fp16")];
            tensor<string, []> var_2744_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2744_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2745_epsilon_0 = const()[name = tensor<string, []>("op_2745_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_2744_cast_fp16_to_fp32 = cast(dtype = var_2744_cast_fp16_to_fp32_dtype_0, x = var_2744_cast_fp16)[name = tensor<string, []>("cast_410")];
            tensor<fp32, [1, 1, 12, 1]> var_2745 = rsqrt(epsilon = var_2745_epsilon_0, x = var_2744_cast_fp16_to_fp32)[name = tensor<string, []>("op_2745")];
            tensor<string, []> var_2745_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2745_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_2745_to_fp16 = cast(dtype = var_2745_to_fp16_dtype_0, x = var_2745)[name = tensor<string, []>("cast_409")];
            tensor<fp16, [1, 1, 12, 128]> x_617_cast_fp16 = mul(x = x_611_cast_fp16, y = var_2745_to_fp16)[name = tensor<string, []>("x_617_cast_fp16")];
            tensor<fp16, [128]> layers_15_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357774848)))];
            tensor<fp16, [1, 1, 12, 128]> var_2747_cast_fp16 = mul(x = layers_15_self_attn_q_norm_weight_to_fp16, y = x_617_cast_fp16)[name = tensor<string, []>("op_2747_cast_fp16")];
            tensor<int32, [4]> q_61_perm_0 = const()[name = tensor<string, []>("q_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357775168)))];
            tensor<fp16, [1, 1, 512]> linear_106_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = x_609_cast_fp16)[name = tensor<string, []>("linear_106_cast_fp16")];
            tensor<int32, [4]> var_2751 = const()[name = tensor<string, []>("op_2751"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_619_cast_fp16 = reshape(shape = var_2751, x = linear_106_cast_fp16)[name = tensor<string, []>("x_619_cast_fp16")];
            tensor<string, []> x_619_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_619_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2705_promoted_2 = const()[name = tensor<string, []>("op_2705_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_619_cast_fp16_to_fp32 = cast(dtype = x_619_cast_fp16_to_fp32_dtype_0, x = x_619_cast_fp16)[name = tensor<string, []>("cast_408")];
            tensor<fp32, [1, 1, 4, 128]> var_2755 = pow(x = x_619_cast_fp16_to_fp32, y = var_2705_promoted_2)[name = tensor<string, []>("op_2755")];
            tensor<int32, [1]> var_125_axes_0 = const()[name = tensor<string, []>("var_125_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_125_keep_dims_0 = const()[name = tensor<string, []>("var_125_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_125 = reduce_mean(axes = var_125_axes_0, keep_dims = var_125_keep_dims_0, x = var_2755)[name = tensor<string, []>("var_125")];
            tensor<string, []> var_125_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_125_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2759_to_fp16 = const()[name = tensor<string, []>("op_2759_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_125_to_fp16 = cast(dtype = var_125_to_fp16_dtype_0, x = var_125)[name = tensor<string, []>("cast_407")];
            tensor<fp16, [1, 1, 4, 1]> var_2760_cast_fp16 = add(x = var_125_to_fp16, y = var_2759_to_fp16)[name = tensor<string, []>("op_2760_cast_fp16")];
            tensor<string, []> var_2760_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2760_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2761_epsilon_0 = const()[name = tensor<string, []>("op_2761_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_2760_cast_fp16_to_fp32 = cast(dtype = var_2760_cast_fp16_to_fp32_dtype_0, x = var_2760_cast_fp16)[name = tensor<string, []>("cast_406")];
            tensor<fp32, [1, 1, 4, 1]> var_2761 = rsqrt(epsilon = var_2761_epsilon_0, x = var_2760_cast_fp16_to_fp32)[name = tensor<string, []>("op_2761")];
            tensor<string, []> var_2761_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2761_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_2761_to_fp16 = cast(dtype = var_2761_to_fp16_dtype_0, x = var_2761)[name = tensor<string, []>("cast_405")];
            tensor<fp16, [1, 1, 4, 128]> x_625_cast_fp16 = mul(x = x_619_cast_fp16, y = var_2761_to_fp16)[name = tensor<string, []>("x_625_cast_fp16")];
            tensor<fp16, [128]> layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358299520)))];
            tensor<fp16, [1, 1, 4, 128]> var_2763_cast_fp16 = mul(x = layers_15_self_attn_k_norm_weight_to_fp16, y = x_625_cast_fp16)[name = tensor<string, []>("op_2763_cast_fp16")];
            tensor<int32, [4]> k_61_perm_0 = const()[name = tensor<string, []>("k_61_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358299840)))];
            tensor<fp16, [1, 1, 512]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = x_609_cast_fp16)[name = tensor<string, []>("linear_107_cast_fp16")];
            tensor<int32, [4]> var_2767 = const()[name = tensor<string, []>("op_2767"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_2768_cast_fp16 = reshape(shape = var_2767, x = linear_107_cast_fp16)[name = tensor<string, []>("op_2768_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = tensor<string, []>("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_61_cast_fp16 = transpose(perm = q_61_perm_0, x = var_2747_cast_fp16)[name = tensor<string, []>("transpose_51")];
            tensor<fp16, [1, 12, 1, 128]> var_2772_cast_fp16 = mul(x = q_61_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2772_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = tensor<string, []>("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = tensor<string, []>("x1_61_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = tensor<string, []>("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_61_cast_fp16)[name = tensor<string, []>("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = tensor<string, []>("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = tensor<string, []>("x2_61_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = tensor<string, []>("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_61_cast_fp16)[name = tensor<string, []>("x2_61_cast_fp16")];
            tensor<fp16, []> const_31_promoted_to_fp16 = const()[name = tensor<string, []>("const_31_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_2775_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_31_promoted_to_fp16)[name = tensor<string, []>("op_2775_cast_fp16")];
            tensor<bool, []> var_2777_interleave_0 = const()[name = tensor<string, []>("op_2777_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_2777_cast_fp16 = concat(axis = var_2706, interleave = var_2777_interleave_0, values = (var_2775_cast_fp16, x1_61_cast_fp16))[name = tensor<string, []>("op_2777_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_2778_cast_fp16 = mul(x = var_2777_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2778_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_63_cast_fp16 = add(x = var_2772_cast_fp16, y = var_2778_cast_fp16)[name = tensor<string, []>("q_63_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_61_cast_fp16 = transpose(perm = k_61_perm_0, x = var_2763_cast_fp16)[name = tensor<string, []>("transpose_50")];
            tensor<fp16, [1, 4, 1, 128]> var_2780_cast_fp16 = mul(x = k_61_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2780_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = tensor<string, []>("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = tensor<string, []>("x1_63_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = tensor<string, []>("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_61_cast_fp16)[name = tensor<string, []>("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = tensor<string, []>("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = tensor<string, []>("x2_63_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = tensor<string, []>("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_61_cast_fp16)[name = tensor<string, []>("x2_63_cast_fp16")];
            tensor<fp16, []> const_32_promoted_to_fp16 = const()[name = tensor<string, []>("const_32_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_2783_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_32_promoted_to_fp16)[name = tensor<string, []>("op_2783_cast_fp16")];
            tensor<bool, []> var_2785_interleave_0 = const()[name = tensor<string, []>("op_2785_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_2785_cast_fp16 = concat(axis = var_2706, interleave = var_2785_interleave_0, values = (var_2783_cast_fp16, x1_63_cast_fp16))[name = tensor<string, []>("op_2785_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_2786_cast_fp16 = mul(x = var_2785_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2786_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_63_cast_fp16 = add(x = var_2780_cast_fp16, y = var_2786_cast_fp16)[name = tensor<string, []>("k_63_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2789_cast_fp16 = mul(x = k_cache_31_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2789_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2790_cast_fp16 = mul(x = k_63_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2790_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_31_cast_fp16 = add(x = var_2789_cast_fp16, y = var_2790_cast_fp16)[name = tensor<string, []>("k_full_31_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2793_cast_fp16 = mul(x = v_cache_31_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2793_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_2768_cast_fp16)[name = tensor<string, []>("transpose_49")];
            tensor<fp16, [1, 4, 2048, 128]> var_2794_cast_fp16 = mul(x = v_31_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2794_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_31_cast_fp16 = add(x = var_2793_cast_fp16, y = var_2794_cast_fp16)[name = tensor<string, []>("v_full_31_cast_fp16")];
            tensor<int32, [1]> var_2796_axes_0 = const()[name = tensor<string, []>("op_2796_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2796_cast_fp16 = expand_dims(axes = var_2796_axes_0, x = k_full_31_cast_fp16)[name = tensor<string, []>("op_2796_cast_fp16")];
            tensor<int32, [5]> var_2798_reps_0 = const()[name = tensor<string, []>("op_2798_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2798_cast_fp16 = tile(reps = var_2798_reps_0, x = var_2796_cast_fp16)[name = tensor<string, []>("op_2798_cast_fp16")];
            tensor<int32, [4]> var_2799 = const()[name = tensor<string, []>("op_2799"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_31_cast_fp16 = reshape(shape = var_2799, x = var_2798_cast_fp16)[name = tensor<string, []>("k_rep_31_cast_fp16")];
            tensor<int32, [1]> var_2801_axes_0 = const()[name = tensor<string, []>("op_2801_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2801_cast_fp16 = expand_dims(axes = var_2801_axes_0, x = v_full_31_cast_fp16)[name = tensor<string, []>("op_2801_cast_fp16")];
            tensor<int32, [5]> var_2803_reps_0 = const()[name = tensor<string, []>("op_2803_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2803_cast_fp16 = tile(reps = var_2803_reps_0, x = var_2801_cast_fp16)[name = tensor<string, []>("op_2803_cast_fp16")];
            tensor<int32, [4]> var_2804 = const()[name = tensor<string, []>("op_2804"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_31_cast_fp16 = reshape(shape = var_2804, x = var_2803_cast_fp16)[name = tensor<string, []>("v_rep_31_cast_fp16")];
            tensor<bool, []> var_2807_transpose_x_1 = const()[name = tensor<string, []>("op_2807_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2807_transpose_y_1 = const()[name = tensor<string, []>("op_2807_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_2807_cast_fp16 = matmul(transpose_x = var_2807_transpose_x_1, transpose_y = var_2807_transpose_y_1, x = q_63_cast_fp16, y = k_rep_31_cast_fp16)[name = tensor<string, []>("op_2807_cast_fp16")];
            tensor<fp16, []> var_2808_to_fp16 = const()[name = tensor<string, []>("op_2808_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_61_cast_fp16 = mul(x = var_2807_cast_fp16, y = var_2808_to_fp16)[name = tensor<string, []>("attn_61_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_63_cast_fp16 = add(x = attn_61_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_63_cast_fp16")];
            tensor<string, []> input_63_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_63_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_63_cast_fp16_to_fp32 = cast(dtype = input_63_cast_fp16_to_fp32_dtype_0, x = input_63_cast_fp16)[name = tensor<string, []>("cast_404")];
            tensor<fp32, [1, 12, 1, 2048]> attn_63 = softmax(axis = var_2706, x = input_63_cast_fp16_to_fp32)[name = tensor<string, []>("attn_63")];
            tensor<bool, []> out_31_transpose_x_0 = const()[name = tensor<string, []>("out_31_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_31_transpose_y_0 = const()[name = tensor<string, []>("out_31_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_63_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_63_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_63_to_fp16 = cast(dtype = attn_63_to_fp16_dtype_0, x = attn_63)[name = tensor<string, []>("cast_403")];
            tensor<fp16, [1, 12, 1, 128]> out_31_cast_fp16 = matmul(transpose_x = out_31_transpose_x_0, transpose_y = out_31_transpose_y_0, x = attn_63_to_fp16, y = v_rep_31_cast_fp16)[name = tensor<string, []>("out_31_cast_fp16")];
            tensor<int32, [4]> var_2813_perm_0 = const()[name = tensor<string, []>("op_2813_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2815 = const()[name = tensor<string, []>("op_2815"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_2813_cast_fp16 = transpose(perm = var_2813_perm_0, x = out_31_cast_fp16)[name = tensor<string, []>("transpose_48")];
            tensor<fp16, [1, 1, 1536]> x_627_cast_fp16 = reshape(shape = var_2815, x = var_2813_cast_fp16)[name = tensor<string, []>("x_627_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(358824192)))];
            tensor<fp16, [1, 1, 512]> linear_108_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = x_627_cast_fp16)[name = tensor<string, []>("linear_108_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_629_cast_fp16 = add(x = x_601_cast_fp16, y = linear_108_cast_fp16)[name = tensor<string, []>("x_629_cast_fp16")];
            tensor<string, []> x_629_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_629_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2705_promoted_3 = const()[name = tensor<string, []>("op_2705_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_629_cast_fp16_to_fp32 = cast(dtype = x_629_cast_fp16_to_fp32_dtype_0, x = x_629_cast_fp16)[name = tensor<string, []>("cast_402")];
            tensor<fp32, [1, 1, 512]> var_2826 = pow(x = x_629_cast_fp16_to_fp32, y = var_2705_promoted_3)[name = tensor<string, []>("op_2826")];
            tensor<int32, [1]> var_127_axes_0 = const()[name = tensor<string, []>("var_127_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_127_keep_dims_0 = const()[name = tensor<string, []>("var_127_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_127 = reduce_mean(axes = var_127_axes_0, keep_dims = var_127_keep_dims_0, x = var_2826)[name = tensor<string, []>("var_127")];
            tensor<string, []> var_127_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_127_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2830_to_fp16 = const()[name = tensor<string, []>("op_2830_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_127_to_fp16 = cast(dtype = var_127_to_fp16_dtype_0, x = var_127)[name = tensor<string, []>("cast_401")];
            tensor<fp16, [1, 1, 1]> var_2831_cast_fp16 = add(x = var_127_to_fp16, y = var_2830_to_fp16)[name = tensor<string, []>("op_2831_cast_fp16")];
            tensor<string, []> var_2831_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2831_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2832_epsilon_0 = const()[name = tensor<string, []>("op_2832_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2831_cast_fp16_to_fp32 = cast(dtype = var_2831_cast_fp16_to_fp32_dtype_0, x = var_2831_cast_fp16)[name = tensor<string, []>("cast_400")];
            tensor<fp32, [1, 1, 1]> var_2832 = rsqrt(epsilon = var_2832_epsilon_0, x = var_2831_cast_fp16_to_fp32)[name = tensor<string, []>("op_2832")];
            tensor<string, []> var_2832_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2832_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2832_to_fp16 = cast(dtype = var_2832_to_fp16_dtype_0, x = var_2832)[name = tensor<string, []>("cast_399")];
            tensor<fp16, [1, 1, 512]> x_635_cast_fp16 = mul(x = x_629_cast_fp16, y = var_2832_to_fp16)[name = tensor<string, []>("x_635_cast_fp16")];
            tensor<fp16, [512]> layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360397120)))];
            tensor<fp16, [1, 1, 512]> x_637_cast_fp16 = mul(x = layers_15_post_attention_layernorm_weight_to_fp16, y = x_635_cast_fp16)[name = tensor<string, []>("x_637_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_15_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360398208)))];
            tensor<fp16, [1, 1, 1536]> linear_109_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_15_mlp_gate_proj_weight_to_fp16, x = x_637_cast_fp16)[name = tensor<string, []>("linear_109_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_2843_cast_fp16 = silu(x = linear_109_cast_fp16)[name = tensor<string, []>("op_2843_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_15_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(361971136)))];
            tensor<fp16, [1, 1, 1536]> linear_110_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_15_mlp_up_proj_weight_to_fp16, x = x_637_cast_fp16)[name = tensor<string, []>("linear_110_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_639_cast_fp16 = mul(x = var_2843_cast_fp16, y = linear_110_cast_fp16)[name = tensor<string, []>("x_639_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_15_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_15_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(363544064)))];
            tensor<fp16, [1, 1, 512]> linear_111_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_15_mlp_down_proj_weight_to_fp16, x = x_639_cast_fp16)[name = tensor<string, []>("linear_111_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_641_cast_fp16 = add(x = x_629_cast_fp16, y = linear_111_cast_fp16)[name = tensor<string, []>("x_641_cast_fp16")];
            tensor<string, []> x_641_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_641_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_33_begin_0 = const()[name = tensor<string, []>("k_cache_33_begin_0"), val = tensor<int32, [5]>([16, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_33_end_0 = const()[name = tensor<string, []>("k_cache_33_end_0"), val = tensor<int32, [5]>([17, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_33_end_mask_0 = const()[name = tensor<string, []>("k_cache_33_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_33_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_33_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_33_cast_fp16 = slice_by_index(begin = k_cache_33_begin_0, end = k_cache_33_end_0, end_mask = k_cache_33_end_mask_0, squeeze_mask = k_cache_33_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_33_cast_fp16")];
            tensor<int32, [5]> v_cache_33_begin_0 = const()[name = tensor<string, []>("v_cache_33_begin_0"), val = tensor<int32, [5]>([16, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_33_end_0 = const()[name = tensor<string, []>("v_cache_33_end_0"), val = tensor<int32, [5]>([17, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_33_end_mask_0 = const()[name = tensor<string, []>("v_cache_33_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_33_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_33_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_33_cast_fp16 = slice_by_index(begin = v_cache_33_begin_0, end = v_cache_33_end_0, end_mask = v_cache_33_end_mask_0, squeeze_mask = v_cache_33_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_33_cast_fp16")];
            tensor<int32, []> var_2875 = const()[name = tensor<string, []>("op_2875"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_2874_promoted = const()[name = tensor<string, []>("op_2874_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_641_cast_fp16_to_fp32 = cast(dtype = x_641_cast_fp16_to_fp32_dtype_0, x = x_641_cast_fp16)[name = tensor<string, []>("cast_398")];
            tensor<fp32, [1, 1, 512]> var_2884 = pow(x = x_641_cast_fp16_to_fp32, y = var_2874_promoted)[name = tensor<string, []>("op_2884")];
            tensor<int32, [1]> var_129_axes_0 = const()[name = tensor<string, []>("var_129_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_129_keep_dims_0 = const()[name = tensor<string, []>("var_129_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_129 = reduce_mean(axes = var_129_axes_0, keep_dims = var_129_keep_dims_0, x = var_2884)[name = tensor<string, []>("var_129")];
            tensor<string, []> var_129_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_129_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2888_to_fp16 = const()[name = tensor<string, []>("op_2888_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_129_to_fp16 = cast(dtype = var_129_to_fp16_dtype_0, x = var_129)[name = tensor<string, []>("cast_397")];
            tensor<fp16, [1, 1, 1]> var_2889_cast_fp16 = add(x = var_129_to_fp16, y = var_2888_to_fp16)[name = tensor<string, []>("op_2889_cast_fp16")];
            tensor<string, []> var_2889_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2889_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2890_epsilon_0 = const()[name = tensor<string, []>("op_2890_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_2889_cast_fp16_to_fp32 = cast(dtype = var_2889_cast_fp16_to_fp32_dtype_0, x = var_2889_cast_fp16)[name = tensor<string, []>("cast_396")];
            tensor<fp32, [1, 1, 1]> var_2890 = rsqrt(epsilon = var_2890_epsilon_0, x = var_2889_cast_fp16_to_fp32)[name = tensor<string, []>("op_2890")];
            tensor<string, []> var_2890_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2890_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_2890_to_fp16 = cast(dtype = var_2890_to_fp16_dtype_0, x = var_2890)[name = tensor<string, []>("cast_395")];
            tensor<fp16, [1, 1, 512]> x_647_cast_fp16 = mul(x = x_641_cast_fp16, y = var_2890_to_fp16)[name = tensor<string, []>("x_647_cast_fp16")];
            tensor<fp16, [512]> layers_16_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(365116992)))];
            tensor<fp16, [1, 1, 512]> x_649_cast_fp16 = mul(x = layers_16_input_layernorm_weight_to_fp16, y = x_647_cast_fp16)[name = tensor<string, []>("x_649_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(365118080)))];
            tensor<fp16, [1, 1, 1536]> linear_112_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = x_649_cast_fp16)[name = tensor<string, []>("linear_112_cast_fp16")];
            tensor<int32, [4]> var_2904 = const()[name = tensor<string, []>("op_2904"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_651_cast_fp16 = reshape(shape = var_2904, x = linear_112_cast_fp16)[name = tensor<string, []>("x_651_cast_fp16")];
            tensor<string, []> x_651_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_651_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2874_promoted_1 = const()[name = tensor<string, []>("op_2874_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_651_cast_fp16_to_fp32 = cast(dtype = x_651_cast_fp16_to_fp32_dtype_0, x = x_651_cast_fp16)[name = tensor<string, []>("cast_394")];
            tensor<fp32, [1, 1, 12, 128]> var_2908 = pow(x = x_651_cast_fp16_to_fp32, y = var_2874_promoted_1)[name = tensor<string, []>("op_2908")];
            tensor<int32, [1]> var_131_axes_0 = const()[name = tensor<string, []>("var_131_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_131_keep_dims_0 = const()[name = tensor<string, []>("var_131_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_131 = reduce_mean(axes = var_131_axes_0, keep_dims = var_131_keep_dims_0, x = var_2908)[name = tensor<string, []>("var_131")];
            tensor<string, []> var_131_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_131_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2912_to_fp16 = const()[name = tensor<string, []>("op_2912_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_131_to_fp16 = cast(dtype = var_131_to_fp16_dtype_0, x = var_131)[name = tensor<string, []>("cast_393")];
            tensor<fp16, [1, 1, 12, 1]> var_2913_cast_fp16 = add(x = var_131_to_fp16, y = var_2912_to_fp16)[name = tensor<string, []>("op_2913_cast_fp16")];
            tensor<string, []> var_2913_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2913_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2914_epsilon_0 = const()[name = tensor<string, []>("op_2914_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_2913_cast_fp16_to_fp32 = cast(dtype = var_2913_cast_fp16_to_fp32_dtype_0, x = var_2913_cast_fp16)[name = tensor<string, []>("cast_392")];
            tensor<fp32, [1, 1, 12, 1]> var_2914 = rsqrt(epsilon = var_2914_epsilon_0, x = var_2913_cast_fp16_to_fp32)[name = tensor<string, []>("op_2914")];
            tensor<string, []> var_2914_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2914_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_2914_to_fp16 = cast(dtype = var_2914_to_fp16_dtype_0, x = var_2914)[name = tensor<string, []>("cast_391")];
            tensor<fp16, [1, 1, 12, 128]> x_657_cast_fp16 = mul(x = x_651_cast_fp16, y = var_2914_to_fp16)[name = tensor<string, []>("x_657_cast_fp16")];
            tensor<fp16, [128]> layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366691008)))];
            tensor<fp16, [1, 1, 12, 128]> var_2916_cast_fp16 = mul(x = layers_16_self_attn_q_norm_weight_to_fp16, y = x_657_cast_fp16)[name = tensor<string, []>("op_2916_cast_fp16")];
            tensor<int32, [4]> q_65_perm_0 = const()[name = tensor<string, []>("q_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(366691328)))];
            tensor<fp16, [1, 1, 512]> linear_113_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = x_649_cast_fp16)[name = tensor<string, []>("linear_113_cast_fp16")];
            tensor<int32, [4]> var_2920 = const()[name = tensor<string, []>("op_2920"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_659_cast_fp16 = reshape(shape = var_2920, x = linear_113_cast_fp16)[name = tensor<string, []>("x_659_cast_fp16")];
            tensor<string, []> x_659_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_659_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2874_promoted_2 = const()[name = tensor<string, []>("op_2874_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_659_cast_fp16_to_fp32 = cast(dtype = x_659_cast_fp16_to_fp32_dtype_0, x = x_659_cast_fp16)[name = tensor<string, []>("cast_390")];
            tensor<fp32, [1, 1, 4, 128]> var_2924 = pow(x = x_659_cast_fp16_to_fp32, y = var_2874_promoted_2)[name = tensor<string, []>("op_2924")];
            tensor<int32, [1]> var_133_axes_0 = const()[name = tensor<string, []>("var_133_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_133_keep_dims_0 = const()[name = tensor<string, []>("var_133_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_133 = reduce_mean(axes = var_133_axes_0, keep_dims = var_133_keep_dims_0, x = var_2924)[name = tensor<string, []>("var_133")];
            tensor<string, []> var_133_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_133_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2928_to_fp16 = const()[name = tensor<string, []>("op_2928_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_133_to_fp16 = cast(dtype = var_133_to_fp16_dtype_0, x = var_133)[name = tensor<string, []>("cast_389")];
            tensor<fp16, [1, 1, 4, 1]> var_2929_cast_fp16 = add(x = var_133_to_fp16, y = var_2928_to_fp16)[name = tensor<string, []>("op_2929_cast_fp16")];
            tensor<string, []> var_2929_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_2929_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2930_epsilon_0 = const()[name = tensor<string, []>("op_2930_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_2929_cast_fp16_to_fp32 = cast(dtype = var_2929_cast_fp16_to_fp32_dtype_0, x = var_2929_cast_fp16)[name = tensor<string, []>("cast_388")];
            tensor<fp32, [1, 1, 4, 1]> var_2930 = rsqrt(epsilon = var_2930_epsilon_0, x = var_2929_cast_fp16_to_fp32)[name = tensor<string, []>("op_2930")];
            tensor<string, []> var_2930_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_2930_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_2930_to_fp16 = cast(dtype = var_2930_to_fp16_dtype_0, x = var_2930)[name = tensor<string, []>("cast_387")];
            tensor<fp16, [1, 1, 4, 128]> x_665_cast_fp16 = mul(x = x_659_cast_fp16, y = var_2930_to_fp16)[name = tensor<string, []>("x_665_cast_fp16")];
            tensor<fp16, [128]> layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367215680)))];
            tensor<fp16, [1, 1, 4, 128]> var_2932_cast_fp16 = mul(x = layers_16_self_attn_k_norm_weight_to_fp16, y = x_665_cast_fp16)[name = tensor<string, []>("op_2932_cast_fp16")];
            tensor<int32, [4]> k_65_perm_0 = const()[name = tensor<string, []>("k_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367216000)))];
            tensor<fp16, [1, 1, 512]> linear_114_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = x_649_cast_fp16)[name = tensor<string, []>("linear_114_cast_fp16")];
            tensor<int32, [4]> var_2936 = const()[name = tensor<string, []>("op_2936"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_2937_cast_fp16 = reshape(shape = var_2936, x = linear_114_cast_fp16)[name = tensor<string, []>("op_2937_cast_fp16")];
            tensor<int32, [4]> v_33_perm_0 = const()[name = tensor<string, []>("v_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_65_cast_fp16 = transpose(perm = q_65_perm_0, x = var_2916_cast_fp16)[name = tensor<string, []>("transpose_47")];
            tensor<fp16, [1, 12, 1, 128]> var_2941_cast_fp16 = mul(x = q_65_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2941_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = tensor<string, []>("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = tensor<string, []>("x1_65_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = tensor<string, []>("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_65_cast_fp16)[name = tensor<string, []>("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = tensor<string, []>("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = tensor<string, []>("x2_65_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = tensor<string, []>("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_65_cast_fp16)[name = tensor<string, []>("x2_65_cast_fp16")];
            tensor<fp16, []> const_33_promoted_to_fp16 = const()[name = tensor<string, []>("const_33_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_2944_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_33_promoted_to_fp16)[name = tensor<string, []>("op_2944_cast_fp16")];
            tensor<bool, []> var_2946_interleave_0 = const()[name = tensor<string, []>("op_2946_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_2946_cast_fp16 = concat(axis = var_2875, interleave = var_2946_interleave_0, values = (var_2944_cast_fp16, x1_65_cast_fp16))[name = tensor<string, []>("op_2946_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_2947_cast_fp16 = mul(x = var_2946_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2947_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_67_cast_fp16 = add(x = var_2941_cast_fp16, y = var_2947_cast_fp16)[name = tensor<string, []>("q_67_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_65_cast_fp16 = transpose(perm = k_65_perm_0, x = var_2932_cast_fp16)[name = tensor<string, []>("transpose_46")];
            tensor<fp16, [1, 4, 1, 128]> var_2949_cast_fp16 = mul(x = k_65_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_2949_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = tensor<string, []>("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = tensor<string, []>("x1_67_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = tensor<string, []>("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_65_cast_fp16)[name = tensor<string, []>("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = tensor<string, []>("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = tensor<string, []>("x2_67_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = tensor<string, []>("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_65_cast_fp16)[name = tensor<string, []>("x2_67_cast_fp16")];
            tensor<fp16, []> const_34_promoted_to_fp16 = const()[name = tensor<string, []>("const_34_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_2952_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_34_promoted_to_fp16)[name = tensor<string, []>("op_2952_cast_fp16")];
            tensor<bool, []> var_2954_interleave_0 = const()[name = tensor<string, []>("op_2954_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_2954_cast_fp16 = concat(axis = var_2875, interleave = var_2954_interleave_0, values = (var_2952_cast_fp16, x1_67_cast_fp16))[name = tensor<string, []>("op_2954_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_2955_cast_fp16 = mul(x = var_2954_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_2955_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_67_cast_fp16 = add(x = var_2949_cast_fp16, y = var_2955_cast_fp16)[name = tensor<string, []>("k_67_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2958_cast_fp16 = mul(x = k_cache_33_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2958_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2959_cast_fp16 = mul(x = k_67_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2959_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_33_cast_fp16 = add(x = var_2958_cast_fp16, y = var_2959_cast_fp16)[name = tensor<string, []>("k_full_33_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_2962_cast_fp16 = mul(x = v_cache_33_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_2962_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_2937_cast_fp16)[name = tensor<string, []>("transpose_45")];
            tensor<fp16, [1, 4, 2048, 128]> var_2963_cast_fp16 = mul(x = v_33_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_2963_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_33_cast_fp16 = add(x = var_2962_cast_fp16, y = var_2963_cast_fp16)[name = tensor<string, []>("v_full_33_cast_fp16")];
            tensor<int32, [1]> var_2965_axes_0 = const()[name = tensor<string, []>("op_2965_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2965_cast_fp16 = expand_dims(axes = var_2965_axes_0, x = k_full_33_cast_fp16)[name = tensor<string, []>("op_2965_cast_fp16")];
            tensor<int32, [5]> var_2967_reps_0 = const()[name = tensor<string, []>("op_2967_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2967_cast_fp16 = tile(reps = var_2967_reps_0, x = var_2965_cast_fp16)[name = tensor<string, []>("op_2967_cast_fp16")];
            tensor<int32, [4]> var_2968 = const()[name = tensor<string, []>("op_2968"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_33_cast_fp16 = reshape(shape = var_2968, x = var_2967_cast_fp16)[name = tensor<string, []>("k_rep_33_cast_fp16")];
            tensor<int32, [1]> var_2970_axes_0 = const()[name = tensor<string, []>("op_2970_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_2970_cast_fp16 = expand_dims(axes = var_2970_axes_0, x = v_full_33_cast_fp16)[name = tensor<string, []>("op_2970_cast_fp16")];
            tensor<int32, [5]> var_2972_reps_0 = const()[name = tensor<string, []>("op_2972_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_2972_cast_fp16 = tile(reps = var_2972_reps_0, x = var_2970_cast_fp16)[name = tensor<string, []>("op_2972_cast_fp16")];
            tensor<int32, [4]> var_2973 = const()[name = tensor<string, []>("op_2973"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_33_cast_fp16 = reshape(shape = var_2973, x = var_2972_cast_fp16)[name = tensor<string, []>("v_rep_33_cast_fp16")];
            tensor<bool, []> var_2976_transpose_x_1 = const()[name = tensor<string, []>("op_2976_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_2976_transpose_y_1 = const()[name = tensor<string, []>("op_2976_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_2976_cast_fp16 = matmul(transpose_x = var_2976_transpose_x_1, transpose_y = var_2976_transpose_y_1, x = q_67_cast_fp16, y = k_rep_33_cast_fp16)[name = tensor<string, []>("op_2976_cast_fp16")];
            tensor<fp16, []> var_2977_to_fp16 = const()[name = tensor<string, []>("op_2977_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_65_cast_fp16 = mul(x = var_2976_cast_fp16, y = var_2977_to_fp16)[name = tensor<string, []>("attn_65_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_67_cast_fp16 = add(x = attn_65_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_67_cast_fp16")];
            tensor<string, []> input_67_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_67_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_67_cast_fp16_to_fp32 = cast(dtype = input_67_cast_fp16_to_fp32_dtype_0, x = input_67_cast_fp16)[name = tensor<string, []>("cast_386")];
            tensor<fp32, [1, 12, 1, 2048]> attn_67 = softmax(axis = var_2875, x = input_67_cast_fp16_to_fp32)[name = tensor<string, []>("attn_67")];
            tensor<bool, []> out_33_transpose_x_0 = const()[name = tensor<string, []>("out_33_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_33_transpose_y_0 = const()[name = tensor<string, []>("out_33_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_67_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_67_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_67_to_fp16 = cast(dtype = attn_67_to_fp16_dtype_0, x = attn_67)[name = tensor<string, []>("cast_385")];
            tensor<fp16, [1, 12, 1, 128]> out_33_cast_fp16 = matmul(transpose_x = out_33_transpose_x_0, transpose_y = out_33_transpose_y_0, x = attn_67_to_fp16, y = v_rep_33_cast_fp16)[name = tensor<string, []>("out_33_cast_fp16")];
            tensor<int32, [4]> var_2982_perm_0 = const()[name = tensor<string, []>("op_2982_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2984 = const()[name = tensor<string, []>("op_2984"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_2982_cast_fp16 = transpose(perm = var_2982_perm_0, x = out_33_cast_fp16)[name = tensor<string, []>("transpose_44")];
            tensor<fp16, [1, 1, 1536]> x_667_cast_fp16 = reshape(shape = var_2984, x = var_2982_cast_fp16)[name = tensor<string, []>("x_667_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(367740352)))];
            tensor<fp16, [1, 1, 512]> linear_115_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = x_667_cast_fp16)[name = tensor<string, []>("linear_115_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_669_cast_fp16 = add(x = x_641_cast_fp16, y = linear_115_cast_fp16)[name = tensor<string, []>("x_669_cast_fp16")];
            tensor<string, []> x_669_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_669_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_2874_promoted_3 = const()[name = tensor<string, []>("op_2874_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_669_cast_fp16_to_fp32 = cast(dtype = x_669_cast_fp16_to_fp32_dtype_0, x = x_669_cast_fp16)[name = tensor<string, []>("cast_384")];
            tensor<fp32, [1, 1, 512]> var_2995 = pow(x = x_669_cast_fp16_to_fp32, y = var_2874_promoted_3)[name = tensor<string, []>("op_2995")];
            tensor<int32, [1]> var_135_axes_0 = const()[name = tensor<string, []>("var_135_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_135_keep_dims_0 = const()[name = tensor<string, []>("var_135_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_135 = reduce_mean(axes = var_135_axes_0, keep_dims = var_135_keep_dims_0, x = var_2995)[name = tensor<string, []>("var_135")];
            tensor<string, []> var_135_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_135_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_2999_to_fp16 = const()[name = tensor<string, []>("op_2999_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_135_to_fp16 = cast(dtype = var_135_to_fp16_dtype_0, x = var_135)[name = tensor<string, []>("cast_383")];
            tensor<fp16, [1, 1, 1]> var_3000_cast_fp16 = add(x = var_135_to_fp16, y = var_2999_to_fp16)[name = tensor<string, []>("op_3000_cast_fp16")];
            tensor<string, []> var_3000_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3000_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3001_epsilon_0 = const()[name = tensor<string, []>("op_3001_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3000_cast_fp16_to_fp32 = cast(dtype = var_3000_cast_fp16_to_fp32_dtype_0, x = var_3000_cast_fp16)[name = tensor<string, []>("cast_382")];
            tensor<fp32, [1, 1, 1]> var_3001 = rsqrt(epsilon = var_3001_epsilon_0, x = var_3000_cast_fp16_to_fp32)[name = tensor<string, []>("op_3001")];
            tensor<string, []> var_3001_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3001_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3001_to_fp16 = cast(dtype = var_3001_to_fp16_dtype_0, x = var_3001)[name = tensor<string, []>("cast_381")];
            tensor<fp16, [1, 1, 512]> x_675_cast_fp16 = mul(x = x_669_cast_fp16, y = var_3001_to_fp16)[name = tensor<string, []>("x_675_cast_fp16")];
            tensor<fp16, [512]> layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369313280)))];
            tensor<fp16, [1, 1, 512]> x_677_cast_fp16 = mul(x = layers_16_post_attention_layernorm_weight_to_fp16, y = x_675_cast_fp16)[name = tensor<string, []>("x_677_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_16_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(369314368)))];
            tensor<fp16, [1, 1, 1536]> linear_116_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_16_mlp_gate_proj_weight_to_fp16, x = x_677_cast_fp16)[name = tensor<string, []>("linear_116_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_3012_cast_fp16 = silu(x = linear_116_cast_fp16)[name = tensor<string, []>("op_3012_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_16_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(370887296)))];
            tensor<fp16, [1, 1, 1536]> linear_117_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_16_mlp_up_proj_weight_to_fp16, x = x_677_cast_fp16)[name = tensor<string, []>("linear_117_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_679_cast_fp16 = mul(x = var_3012_cast_fp16, y = linear_117_cast_fp16)[name = tensor<string, []>("x_679_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_16_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_16_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372460224)))];
            tensor<fp16, [1, 1, 512]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_16_mlp_down_proj_weight_to_fp16, x = x_679_cast_fp16)[name = tensor<string, []>("linear_118_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_681_cast_fp16 = add(x = x_669_cast_fp16, y = linear_118_cast_fp16)[name = tensor<string, []>("x_681_cast_fp16")];
            tensor<string, []> x_681_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_681_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_35_begin_0 = const()[name = tensor<string, []>("k_cache_35_begin_0"), val = tensor<int32, [5]>([17, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_35_end_0 = const()[name = tensor<string, []>("k_cache_35_end_0"), val = tensor<int32, [5]>([18, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_35_end_mask_0 = const()[name = tensor<string, []>("k_cache_35_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_35_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_35_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_35_cast_fp16 = slice_by_index(begin = k_cache_35_begin_0, end = k_cache_35_end_0, end_mask = k_cache_35_end_mask_0, squeeze_mask = k_cache_35_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_35_cast_fp16")];
            tensor<int32, [5]> v_cache_35_begin_0 = const()[name = tensor<string, []>("v_cache_35_begin_0"), val = tensor<int32, [5]>([17, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_35_end_0 = const()[name = tensor<string, []>("v_cache_35_end_0"), val = tensor<int32, [5]>([18, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_35_end_mask_0 = const()[name = tensor<string, []>("v_cache_35_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_35_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_35_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_35_cast_fp16 = slice_by_index(begin = v_cache_35_begin_0, end = v_cache_35_end_0, end_mask = v_cache_35_end_mask_0, squeeze_mask = v_cache_35_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_35_cast_fp16")];
            tensor<int32, []> var_3044 = const()[name = tensor<string, []>("op_3044"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_3043_promoted = const()[name = tensor<string, []>("op_3043_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_681_cast_fp16_to_fp32 = cast(dtype = x_681_cast_fp16_to_fp32_dtype_0, x = x_681_cast_fp16)[name = tensor<string, []>("cast_380")];
            tensor<fp32, [1, 1, 512]> var_3053 = pow(x = x_681_cast_fp16_to_fp32, y = var_3043_promoted)[name = tensor<string, []>("op_3053")];
            tensor<int32, [1]> var_137_axes_0 = const()[name = tensor<string, []>("var_137_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_137_keep_dims_0 = const()[name = tensor<string, []>("var_137_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_137 = reduce_mean(axes = var_137_axes_0, keep_dims = var_137_keep_dims_0, x = var_3053)[name = tensor<string, []>("var_137")];
            tensor<string, []> var_137_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_137_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3057_to_fp16 = const()[name = tensor<string, []>("op_3057_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_137_to_fp16 = cast(dtype = var_137_to_fp16_dtype_0, x = var_137)[name = tensor<string, []>("cast_379")];
            tensor<fp16, [1, 1, 1]> var_3058_cast_fp16 = add(x = var_137_to_fp16, y = var_3057_to_fp16)[name = tensor<string, []>("op_3058_cast_fp16")];
            tensor<string, []> var_3058_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3058_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3059_epsilon_0 = const()[name = tensor<string, []>("op_3059_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3058_cast_fp16_to_fp32 = cast(dtype = var_3058_cast_fp16_to_fp32_dtype_0, x = var_3058_cast_fp16)[name = tensor<string, []>("cast_378")];
            tensor<fp32, [1, 1, 1]> var_3059 = rsqrt(epsilon = var_3059_epsilon_0, x = var_3058_cast_fp16_to_fp32)[name = tensor<string, []>("op_3059")];
            tensor<string, []> var_3059_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3059_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3059_to_fp16 = cast(dtype = var_3059_to_fp16_dtype_0, x = var_3059)[name = tensor<string, []>("cast_377")];
            tensor<fp16, [1, 1, 512]> x_687_cast_fp16 = mul(x = x_681_cast_fp16, y = var_3059_to_fp16)[name = tensor<string, []>("x_687_cast_fp16")];
            tensor<fp16, [512]> layers_17_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374033152)))];
            tensor<fp16, [1, 1, 512]> x_689_cast_fp16 = mul(x = layers_17_input_layernorm_weight_to_fp16, y = x_687_cast_fp16)[name = tensor<string, []>("x_689_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(374034240)))];
            tensor<fp16, [1, 1, 1536]> linear_119_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = x_689_cast_fp16)[name = tensor<string, []>("linear_119_cast_fp16")];
            tensor<int32, [4]> var_3073 = const()[name = tensor<string, []>("op_3073"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_691_cast_fp16 = reshape(shape = var_3073, x = linear_119_cast_fp16)[name = tensor<string, []>("x_691_cast_fp16")];
            tensor<string, []> x_691_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_691_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3043_promoted_1 = const()[name = tensor<string, []>("op_3043_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_691_cast_fp16_to_fp32 = cast(dtype = x_691_cast_fp16_to_fp32_dtype_0, x = x_691_cast_fp16)[name = tensor<string, []>("cast_376")];
            tensor<fp32, [1, 1, 12, 128]> var_3077 = pow(x = x_691_cast_fp16_to_fp32, y = var_3043_promoted_1)[name = tensor<string, []>("op_3077")];
            tensor<int32, [1]> var_139_axes_0 = const()[name = tensor<string, []>("var_139_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_139_keep_dims_0 = const()[name = tensor<string, []>("var_139_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_139 = reduce_mean(axes = var_139_axes_0, keep_dims = var_139_keep_dims_0, x = var_3077)[name = tensor<string, []>("var_139")];
            tensor<string, []> var_139_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_139_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3081_to_fp16 = const()[name = tensor<string, []>("op_3081_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_139_to_fp16 = cast(dtype = var_139_to_fp16_dtype_0, x = var_139)[name = tensor<string, []>("cast_375")];
            tensor<fp16, [1, 1, 12, 1]> var_3082_cast_fp16 = add(x = var_139_to_fp16, y = var_3081_to_fp16)[name = tensor<string, []>("op_3082_cast_fp16")];
            tensor<string, []> var_3082_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3082_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3083_epsilon_0 = const()[name = tensor<string, []>("op_3083_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_3082_cast_fp16_to_fp32 = cast(dtype = var_3082_cast_fp16_to_fp32_dtype_0, x = var_3082_cast_fp16)[name = tensor<string, []>("cast_374")];
            tensor<fp32, [1, 1, 12, 1]> var_3083 = rsqrt(epsilon = var_3083_epsilon_0, x = var_3082_cast_fp16_to_fp32)[name = tensor<string, []>("op_3083")];
            tensor<string, []> var_3083_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3083_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_3083_to_fp16 = cast(dtype = var_3083_to_fp16_dtype_0, x = var_3083)[name = tensor<string, []>("cast_373")];
            tensor<fp16, [1, 1, 12, 128]> x_697_cast_fp16 = mul(x = x_691_cast_fp16, y = var_3083_to_fp16)[name = tensor<string, []>("x_697_cast_fp16")];
            tensor<fp16, [128]> layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375607168)))];
            tensor<fp16, [1, 1, 12, 128]> var_3085_cast_fp16 = mul(x = layers_17_self_attn_q_norm_weight_to_fp16, y = x_697_cast_fp16)[name = tensor<string, []>("op_3085_cast_fp16")];
            tensor<int32, [4]> q_69_perm_0 = const()[name = tensor<string, []>("q_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(375607488)))];
            tensor<fp16, [1, 1, 512]> linear_120_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = x_689_cast_fp16)[name = tensor<string, []>("linear_120_cast_fp16")];
            tensor<int32, [4]> var_3089 = const()[name = tensor<string, []>("op_3089"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_699_cast_fp16 = reshape(shape = var_3089, x = linear_120_cast_fp16)[name = tensor<string, []>("x_699_cast_fp16")];
            tensor<string, []> x_699_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_699_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3043_promoted_2 = const()[name = tensor<string, []>("op_3043_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_699_cast_fp16_to_fp32 = cast(dtype = x_699_cast_fp16_to_fp32_dtype_0, x = x_699_cast_fp16)[name = tensor<string, []>("cast_372")];
            tensor<fp32, [1, 1, 4, 128]> var_3093 = pow(x = x_699_cast_fp16_to_fp32, y = var_3043_promoted_2)[name = tensor<string, []>("op_3093")];
            tensor<int32, [1]> var_141_axes_0 = const()[name = tensor<string, []>("var_141_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_141_keep_dims_0 = const()[name = tensor<string, []>("var_141_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_141 = reduce_mean(axes = var_141_axes_0, keep_dims = var_141_keep_dims_0, x = var_3093)[name = tensor<string, []>("var_141")];
            tensor<string, []> var_141_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_141_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3097_to_fp16 = const()[name = tensor<string, []>("op_3097_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_141_to_fp16 = cast(dtype = var_141_to_fp16_dtype_0, x = var_141)[name = tensor<string, []>("cast_371")];
            tensor<fp16, [1, 1, 4, 1]> var_3098_cast_fp16 = add(x = var_141_to_fp16, y = var_3097_to_fp16)[name = tensor<string, []>("op_3098_cast_fp16")];
            tensor<string, []> var_3098_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3098_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3099_epsilon_0 = const()[name = tensor<string, []>("op_3099_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_3098_cast_fp16_to_fp32 = cast(dtype = var_3098_cast_fp16_to_fp32_dtype_0, x = var_3098_cast_fp16)[name = tensor<string, []>("cast_370")];
            tensor<fp32, [1, 1, 4, 1]> var_3099 = rsqrt(epsilon = var_3099_epsilon_0, x = var_3098_cast_fp16_to_fp32)[name = tensor<string, []>("op_3099")];
            tensor<string, []> var_3099_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3099_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_3099_to_fp16 = cast(dtype = var_3099_to_fp16_dtype_0, x = var_3099)[name = tensor<string, []>("cast_369")];
            tensor<fp16, [1, 1, 4, 128]> x_705_cast_fp16 = mul(x = x_699_cast_fp16, y = var_3099_to_fp16)[name = tensor<string, []>("x_705_cast_fp16")];
            tensor<fp16, [128]> layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(376131840)))];
            tensor<fp16, [1, 1, 4, 128]> var_3101_cast_fp16 = mul(x = layers_17_self_attn_k_norm_weight_to_fp16, y = x_705_cast_fp16)[name = tensor<string, []>("op_3101_cast_fp16")];
            tensor<int32, [4]> k_69_perm_0 = const()[name = tensor<string, []>("k_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(376132160)))];
            tensor<fp16, [1, 1, 512]> linear_121_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = x_689_cast_fp16)[name = tensor<string, []>("linear_121_cast_fp16")];
            tensor<int32, [4]> var_3105 = const()[name = tensor<string, []>("op_3105"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_3106_cast_fp16 = reshape(shape = var_3105, x = linear_121_cast_fp16)[name = tensor<string, []>("op_3106_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = tensor<string, []>("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_69_cast_fp16 = transpose(perm = q_69_perm_0, x = var_3085_cast_fp16)[name = tensor<string, []>("transpose_43")];
            tensor<fp16, [1, 12, 1, 128]> var_3110_cast_fp16 = mul(x = q_69_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3110_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = tensor<string, []>("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = tensor<string, []>("x1_69_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = tensor<string, []>("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_69_cast_fp16)[name = tensor<string, []>("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = tensor<string, []>("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = tensor<string, []>("x2_69_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = tensor<string, []>("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_69_cast_fp16)[name = tensor<string, []>("x2_69_cast_fp16")];
            tensor<fp16, []> const_35_promoted_to_fp16 = const()[name = tensor<string, []>("const_35_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_3113_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_35_promoted_to_fp16)[name = tensor<string, []>("op_3113_cast_fp16")];
            tensor<bool, []> var_3115_interleave_0 = const()[name = tensor<string, []>("op_3115_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_3115_cast_fp16 = concat(axis = var_3044, interleave = var_3115_interleave_0, values = (var_3113_cast_fp16, x1_69_cast_fp16))[name = tensor<string, []>("op_3115_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_3116_cast_fp16 = mul(x = var_3115_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3116_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_71_cast_fp16 = add(x = var_3110_cast_fp16, y = var_3116_cast_fp16)[name = tensor<string, []>("q_71_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_69_cast_fp16 = transpose(perm = k_69_perm_0, x = var_3101_cast_fp16)[name = tensor<string, []>("transpose_42")];
            tensor<fp16, [1, 4, 1, 128]> var_3118_cast_fp16 = mul(x = k_69_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3118_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = tensor<string, []>("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = tensor<string, []>("x1_71_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = tensor<string, []>("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_69_cast_fp16)[name = tensor<string, []>("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = tensor<string, []>("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = tensor<string, []>("x2_71_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = tensor<string, []>("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_69_cast_fp16)[name = tensor<string, []>("x2_71_cast_fp16")];
            tensor<fp16, []> const_36_promoted_to_fp16 = const()[name = tensor<string, []>("const_36_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_3121_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_36_promoted_to_fp16)[name = tensor<string, []>("op_3121_cast_fp16")];
            tensor<bool, []> var_3123_interleave_0 = const()[name = tensor<string, []>("op_3123_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_3123_cast_fp16 = concat(axis = var_3044, interleave = var_3123_interleave_0, values = (var_3121_cast_fp16, x1_71_cast_fp16))[name = tensor<string, []>("op_3123_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_3124_cast_fp16 = mul(x = var_3123_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3124_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_71_cast_fp16 = add(x = var_3118_cast_fp16, y = var_3124_cast_fp16)[name = tensor<string, []>("k_71_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3127_cast_fp16 = mul(x = k_cache_35_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3127_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3128_cast_fp16 = mul(x = k_71_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3128_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_35_cast_fp16 = add(x = var_3127_cast_fp16, y = var_3128_cast_fp16)[name = tensor<string, []>("k_full_35_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3131_cast_fp16 = mul(x = v_cache_35_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3131_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_3106_cast_fp16)[name = tensor<string, []>("transpose_41")];
            tensor<fp16, [1, 4, 2048, 128]> var_3132_cast_fp16 = mul(x = v_35_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3132_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_35_cast_fp16 = add(x = var_3131_cast_fp16, y = var_3132_cast_fp16)[name = tensor<string, []>("v_full_35_cast_fp16")];
            tensor<int32, [1]> var_3134_axes_0 = const()[name = tensor<string, []>("op_3134_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3134_cast_fp16 = expand_dims(axes = var_3134_axes_0, x = k_full_35_cast_fp16)[name = tensor<string, []>("op_3134_cast_fp16")];
            tensor<int32, [5]> var_3136_reps_0 = const()[name = tensor<string, []>("op_3136_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3136_cast_fp16 = tile(reps = var_3136_reps_0, x = var_3134_cast_fp16)[name = tensor<string, []>("op_3136_cast_fp16")];
            tensor<int32, [4]> var_3137 = const()[name = tensor<string, []>("op_3137"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_35_cast_fp16 = reshape(shape = var_3137, x = var_3136_cast_fp16)[name = tensor<string, []>("k_rep_35_cast_fp16")];
            tensor<int32, [1]> var_3139_axes_0 = const()[name = tensor<string, []>("op_3139_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3139_cast_fp16 = expand_dims(axes = var_3139_axes_0, x = v_full_35_cast_fp16)[name = tensor<string, []>("op_3139_cast_fp16")];
            tensor<int32, [5]> var_3141_reps_0 = const()[name = tensor<string, []>("op_3141_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3141_cast_fp16 = tile(reps = var_3141_reps_0, x = var_3139_cast_fp16)[name = tensor<string, []>("op_3141_cast_fp16")];
            tensor<int32, [4]> var_3142 = const()[name = tensor<string, []>("op_3142"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_35_cast_fp16 = reshape(shape = var_3142, x = var_3141_cast_fp16)[name = tensor<string, []>("v_rep_35_cast_fp16")];
            tensor<bool, []> var_3145_transpose_x_1 = const()[name = tensor<string, []>("op_3145_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3145_transpose_y_1 = const()[name = tensor<string, []>("op_3145_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_3145_cast_fp16 = matmul(transpose_x = var_3145_transpose_x_1, transpose_y = var_3145_transpose_y_1, x = q_71_cast_fp16, y = k_rep_35_cast_fp16)[name = tensor<string, []>("op_3145_cast_fp16")];
            tensor<fp16, []> var_3146_to_fp16 = const()[name = tensor<string, []>("op_3146_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_69_cast_fp16 = mul(x = var_3145_cast_fp16, y = var_3146_to_fp16)[name = tensor<string, []>("attn_69_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_71_cast_fp16 = add(x = attn_69_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_71_cast_fp16")];
            tensor<string, []> input_71_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_71_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_71_cast_fp16_to_fp32 = cast(dtype = input_71_cast_fp16_to_fp32_dtype_0, x = input_71_cast_fp16)[name = tensor<string, []>("cast_368")];
            tensor<fp32, [1, 12, 1, 2048]> attn_71 = softmax(axis = var_3044, x = input_71_cast_fp16_to_fp32)[name = tensor<string, []>("attn_71")];
            tensor<bool, []> out_35_transpose_x_0 = const()[name = tensor<string, []>("out_35_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_35_transpose_y_0 = const()[name = tensor<string, []>("out_35_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_71_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_71_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_71_to_fp16 = cast(dtype = attn_71_to_fp16_dtype_0, x = attn_71)[name = tensor<string, []>("cast_367")];
            tensor<fp16, [1, 12, 1, 128]> out_35_cast_fp16 = matmul(transpose_x = out_35_transpose_x_0, transpose_y = out_35_transpose_y_0, x = attn_71_to_fp16, y = v_rep_35_cast_fp16)[name = tensor<string, []>("out_35_cast_fp16")];
            tensor<int32, [4]> var_3151_perm_0 = const()[name = tensor<string, []>("op_3151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3153 = const()[name = tensor<string, []>("op_3153"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_3151_cast_fp16 = transpose(perm = var_3151_perm_0, x = out_35_cast_fp16)[name = tensor<string, []>("transpose_40")];
            tensor<fp16, [1, 1, 1536]> x_707_cast_fp16 = reshape(shape = var_3153, x = var_3151_cast_fp16)[name = tensor<string, []>("x_707_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(376656512)))];
            tensor<fp16, [1, 1, 512]> linear_122_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = x_707_cast_fp16)[name = tensor<string, []>("linear_122_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_709_cast_fp16 = add(x = x_681_cast_fp16, y = linear_122_cast_fp16)[name = tensor<string, []>("x_709_cast_fp16")];
            tensor<string, []> x_709_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_709_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3043_promoted_3 = const()[name = tensor<string, []>("op_3043_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_709_cast_fp16_to_fp32 = cast(dtype = x_709_cast_fp16_to_fp32_dtype_0, x = x_709_cast_fp16)[name = tensor<string, []>("cast_366")];
            tensor<fp32, [1, 1, 512]> var_3164 = pow(x = x_709_cast_fp16_to_fp32, y = var_3043_promoted_3)[name = tensor<string, []>("op_3164")];
            tensor<int32, [1]> var_143_axes_0 = const()[name = tensor<string, []>("var_143_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_143_keep_dims_0 = const()[name = tensor<string, []>("var_143_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_143 = reduce_mean(axes = var_143_axes_0, keep_dims = var_143_keep_dims_0, x = var_3164)[name = tensor<string, []>("var_143")];
            tensor<string, []> var_143_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_143_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3168_to_fp16 = const()[name = tensor<string, []>("op_3168_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_143_to_fp16 = cast(dtype = var_143_to_fp16_dtype_0, x = var_143)[name = tensor<string, []>("cast_365")];
            tensor<fp16, [1, 1, 1]> var_3169_cast_fp16 = add(x = var_143_to_fp16, y = var_3168_to_fp16)[name = tensor<string, []>("op_3169_cast_fp16")];
            tensor<string, []> var_3169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3169_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3170_epsilon_0 = const()[name = tensor<string, []>("op_3170_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3169_cast_fp16_to_fp32 = cast(dtype = var_3169_cast_fp16_to_fp32_dtype_0, x = var_3169_cast_fp16)[name = tensor<string, []>("cast_364")];
            tensor<fp32, [1, 1, 1]> var_3170 = rsqrt(epsilon = var_3170_epsilon_0, x = var_3169_cast_fp16_to_fp32)[name = tensor<string, []>("op_3170")];
            tensor<string, []> var_3170_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3170_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3170_to_fp16 = cast(dtype = var_3170_to_fp16_dtype_0, x = var_3170)[name = tensor<string, []>("cast_363")];
            tensor<fp16, [1, 1, 512]> x_715_cast_fp16 = mul(x = x_709_cast_fp16, y = var_3170_to_fp16)[name = tensor<string, []>("x_715_cast_fp16")];
            tensor<fp16, [512]> layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378229440)))];
            tensor<fp16, [1, 1, 512]> x_717_cast_fp16 = mul(x = layers_17_post_attention_layernorm_weight_to_fp16, y = x_715_cast_fp16)[name = tensor<string, []>("x_717_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_17_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378230528)))];
            tensor<fp16, [1, 1, 1536]> linear_123_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_17_mlp_gate_proj_weight_to_fp16, x = x_717_cast_fp16)[name = tensor<string, []>("linear_123_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_3181_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor<string, []>("op_3181_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_17_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379803456)))];
            tensor<fp16, [1, 1, 1536]> linear_124_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_17_mlp_up_proj_weight_to_fp16, x = x_717_cast_fp16)[name = tensor<string, []>("linear_124_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_719_cast_fp16 = mul(x = var_3181_cast_fp16, y = linear_124_cast_fp16)[name = tensor<string, []>("x_719_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_17_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_17_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381376384)))];
            tensor<fp16, [1, 1, 512]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_17_mlp_down_proj_weight_to_fp16, x = x_719_cast_fp16)[name = tensor<string, []>("linear_125_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_721_cast_fp16 = add(x = x_709_cast_fp16, y = linear_125_cast_fp16)[name = tensor<string, []>("x_721_cast_fp16")];
            tensor<string, []> x_721_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_721_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_37_begin_0 = const()[name = tensor<string, []>("k_cache_37_begin_0"), val = tensor<int32, [5]>([18, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_37_end_0 = const()[name = tensor<string, []>("k_cache_37_end_0"), val = tensor<int32, [5]>([19, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_37_end_mask_0 = const()[name = tensor<string, []>("k_cache_37_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_37_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_37_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_37_cast_fp16 = slice_by_index(begin = k_cache_37_begin_0, end = k_cache_37_end_0, end_mask = k_cache_37_end_mask_0, squeeze_mask = k_cache_37_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_37_cast_fp16")];
            tensor<int32, [5]> v_cache_37_begin_0 = const()[name = tensor<string, []>("v_cache_37_begin_0"), val = tensor<int32, [5]>([18, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_37_end_0 = const()[name = tensor<string, []>("v_cache_37_end_0"), val = tensor<int32, [5]>([19, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_37_end_mask_0 = const()[name = tensor<string, []>("v_cache_37_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_37_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_37_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_37_cast_fp16 = slice_by_index(begin = v_cache_37_begin_0, end = v_cache_37_end_0, end_mask = v_cache_37_end_mask_0, squeeze_mask = v_cache_37_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_37_cast_fp16")];
            tensor<int32, []> var_3213 = const()[name = tensor<string, []>("op_3213"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_3212_promoted = const()[name = tensor<string, []>("op_3212_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_721_cast_fp16_to_fp32 = cast(dtype = x_721_cast_fp16_to_fp32_dtype_0, x = x_721_cast_fp16)[name = tensor<string, []>("cast_362")];
            tensor<fp32, [1, 1, 512]> var_3222 = pow(x = x_721_cast_fp16_to_fp32, y = var_3212_promoted)[name = tensor<string, []>("op_3222")];
            tensor<int32, [1]> var_145_axes_0 = const()[name = tensor<string, []>("var_145_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_145_keep_dims_0 = const()[name = tensor<string, []>("var_145_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_145 = reduce_mean(axes = var_145_axes_0, keep_dims = var_145_keep_dims_0, x = var_3222)[name = tensor<string, []>("var_145")];
            tensor<string, []> var_145_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_145_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3226_to_fp16 = const()[name = tensor<string, []>("op_3226_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_145_to_fp16 = cast(dtype = var_145_to_fp16_dtype_0, x = var_145)[name = tensor<string, []>("cast_361")];
            tensor<fp16, [1, 1, 1]> var_3227_cast_fp16 = add(x = var_145_to_fp16, y = var_3226_to_fp16)[name = tensor<string, []>("op_3227_cast_fp16")];
            tensor<string, []> var_3227_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3227_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3228_epsilon_0 = const()[name = tensor<string, []>("op_3228_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3227_cast_fp16_to_fp32 = cast(dtype = var_3227_cast_fp16_to_fp32_dtype_0, x = var_3227_cast_fp16)[name = tensor<string, []>("cast_360")];
            tensor<fp32, [1, 1, 1]> var_3228 = rsqrt(epsilon = var_3228_epsilon_0, x = var_3227_cast_fp16_to_fp32)[name = tensor<string, []>("op_3228")];
            tensor<string, []> var_3228_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3228_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3228_to_fp16 = cast(dtype = var_3228_to_fp16_dtype_0, x = var_3228)[name = tensor<string, []>("cast_359")];
            tensor<fp16, [1, 1, 512]> x_727_cast_fp16 = mul(x = x_721_cast_fp16, y = var_3228_to_fp16)[name = tensor<string, []>("x_727_cast_fp16")];
            tensor<fp16, [512]> layers_18_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382949312)))];
            tensor<fp16, [1, 1, 512]> x_729_cast_fp16 = mul(x = layers_18_input_layernorm_weight_to_fp16, y = x_727_cast_fp16)[name = tensor<string, []>("x_729_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(382950400)))];
            tensor<fp16, [1, 1, 1536]> linear_126_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = x_729_cast_fp16)[name = tensor<string, []>("linear_126_cast_fp16")];
            tensor<int32, [4]> var_3242 = const()[name = tensor<string, []>("op_3242"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_731_cast_fp16 = reshape(shape = var_3242, x = linear_126_cast_fp16)[name = tensor<string, []>("x_731_cast_fp16")];
            tensor<string, []> x_731_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_731_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3212_promoted_1 = const()[name = tensor<string, []>("op_3212_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_731_cast_fp16_to_fp32 = cast(dtype = x_731_cast_fp16_to_fp32_dtype_0, x = x_731_cast_fp16)[name = tensor<string, []>("cast_358")];
            tensor<fp32, [1, 1, 12, 128]> var_3246 = pow(x = x_731_cast_fp16_to_fp32, y = var_3212_promoted_1)[name = tensor<string, []>("op_3246")];
            tensor<int32, [1]> var_147_axes_0 = const()[name = tensor<string, []>("var_147_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_147_keep_dims_0 = const()[name = tensor<string, []>("var_147_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_147 = reduce_mean(axes = var_147_axes_0, keep_dims = var_147_keep_dims_0, x = var_3246)[name = tensor<string, []>("var_147")];
            tensor<string, []> var_147_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_147_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3250_to_fp16 = const()[name = tensor<string, []>("op_3250_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_147_to_fp16 = cast(dtype = var_147_to_fp16_dtype_0, x = var_147)[name = tensor<string, []>("cast_357")];
            tensor<fp16, [1, 1, 12, 1]> var_3251_cast_fp16 = add(x = var_147_to_fp16, y = var_3250_to_fp16)[name = tensor<string, []>("op_3251_cast_fp16")];
            tensor<string, []> var_3251_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3251_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3252_epsilon_0 = const()[name = tensor<string, []>("op_3252_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_3251_cast_fp16_to_fp32 = cast(dtype = var_3251_cast_fp16_to_fp32_dtype_0, x = var_3251_cast_fp16)[name = tensor<string, []>("cast_356")];
            tensor<fp32, [1, 1, 12, 1]> var_3252 = rsqrt(epsilon = var_3252_epsilon_0, x = var_3251_cast_fp16_to_fp32)[name = tensor<string, []>("op_3252")];
            tensor<string, []> var_3252_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3252_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_3252_to_fp16 = cast(dtype = var_3252_to_fp16_dtype_0, x = var_3252)[name = tensor<string, []>("cast_355")];
            tensor<fp16, [1, 1, 12, 128]> x_737_cast_fp16 = mul(x = x_731_cast_fp16, y = var_3252_to_fp16)[name = tensor<string, []>("x_737_cast_fp16")];
            tensor<fp16, [128]> layers_18_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384523328)))];
            tensor<fp16, [1, 1, 12, 128]> var_3254_cast_fp16 = mul(x = layers_18_self_attn_q_norm_weight_to_fp16, y = x_737_cast_fp16)[name = tensor<string, []>("op_3254_cast_fp16")];
            tensor<int32, [4]> q_73_perm_0 = const()[name = tensor<string, []>("q_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384523648)))];
            tensor<fp16, [1, 1, 512]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = x_729_cast_fp16)[name = tensor<string, []>("linear_127_cast_fp16")];
            tensor<int32, [4]> var_3258 = const()[name = tensor<string, []>("op_3258"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_739_cast_fp16 = reshape(shape = var_3258, x = linear_127_cast_fp16)[name = tensor<string, []>("x_739_cast_fp16")];
            tensor<string, []> x_739_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_739_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3212_promoted_2 = const()[name = tensor<string, []>("op_3212_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_739_cast_fp16_to_fp32 = cast(dtype = x_739_cast_fp16_to_fp32_dtype_0, x = x_739_cast_fp16)[name = tensor<string, []>("cast_354")];
            tensor<fp32, [1, 1, 4, 128]> var_3262 = pow(x = x_739_cast_fp16_to_fp32, y = var_3212_promoted_2)[name = tensor<string, []>("op_3262")];
            tensor<int32, [1]> var_149_axes_0 = const()[name = tensor<string, []>("var_149_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_149_keep_dims_0 = const()[name = tensor<string, []>("var_149_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_149 = reduce_mean(axes = var_149_axes_0, keep_dims = var_149_keep_dims_0, x = var_3262)[name = tensor<string, []>("var_149")];
            tensor<string, []> var_149_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_149_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3266_to_fp16 = const()[name = tensor<string, []>("op_3266_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_149_to_fp16 = cast(dtype = var_149_to_fp16_dtype_0, x = var_149)[name = tensor<string, []>("cast_353")];
            tensor<fp16, [1, 1, 4, 1]> var_3267_cast_fp16 = add(x = var_149_to_fp16, y = var_3266_to_fp16)[name = tensor<string, []>("op_3267_cast_fp16")];
            tensor<string, []> var_3267_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3267_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3268_epsilon_0 = const()[name = tensor<string, []>("op_3268_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_3267_cast_fp16_to_fp32 = cast(dtype = var_3267_cast_fp16_to_fp32_dtype_0, x = var_3267_cast_fp16)[name = tensor<string, []>("cast_352")];
            tensor<fp32, [1, 1, 4, 1]> var_3268 = rsqrt(epsilon = var_3268_epsilon_0, x = var_3267_cast_fp16_to_fp32)[name = tensor<string, []>("op_3268")];
            tensor<string, []> var_3268_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3268_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_3268_to_fp16 = cast(dtype = var_3268_to_fp16_dtype_0, x = var_3268)[name = tensor<string, []>("cast_351")];
            tensor<fp16, [1, 1, 4, 128]> x_745_cast_fp16 = mul(x = x_739_cast_fp16, y = var_3268_to_fp16)[name = tensor<string, []>("x_745_cast_fp16")];
            tensor<fp16, [128]> layers_18_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385048000)))];
            tensor<fp16, [1, 1, 4, 128]> var_3270_cast_fp16 = mul(x = layers_18_self_attn_k_norm_weight_to_fp16, y = x_745_cast_fp16)[name = tensor<string, []>("op_3270_cast_fp16")];
            tensor<int32, [4]> k_73_perm_0 = const()[name = tensor<string, []>("k_73_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385048320)))];
            tensor<fp16, [1, 1, 512]> linear_128_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = x_729_cast_fp16)[name = tensor<string, []>("linear_128_cast_fp16")];
            tensor<int32, [4]> var_3274 = const()[name = tensor<string, []>("op_3274"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_3275_cast_fp16 = reshape(shape = var_3274, x = linear_128_cast_fp16)[name = tensor<string, []>("op_3275_cast_fp16")];
            tensor<int32, [4]> v_37_perm_0 = const()[name = tensor<string, []>("v_37_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_73_cast_fp16 = transpose(perm = q_73_perm_0, x = var_3254_cast_fp16)[name = tensor<string, []>("transpose_39")];
            tensor<fp16, [1, 12, 1, 128]> var_3279_cast_fp16 = mul(x = q_73_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3279_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = tensor<string, []>("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = tensor<string, []>("x1_73_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = tensor<string, []>("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_73_cast_fp16)[name = tensor<string, []>("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = tensor<string, []>("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = tensor<string, []>("x2_73_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = tensor<string, []>("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_73_cast_fp16)[name = tensor<string, []>("x2_73_cast_fp16")];
            tensor<fp16, []> const_37_promoted_to_fp16 = const()[name = tensor<string, []>("const_37_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_3282_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_37_promoted_to_fp16)[name = tensor<string, []>("op_3282_cast_fp16")];
            tensor<bool, []> var_3284_interleave_0 = const()[name = tensor<string, []>("op_3284_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_3284_cast_fp16 = concat(axis = var_3213, interleave = var_3284_interleave_0, values = (var_3282_cast_fp16, x1_73_cast_fp16))[name = tensor<string, []>("op_3284_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_3285_cast_fp16 = mul(x = var_3284_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3285_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_75_cast_fp16 = add(x = var_3279_cast_fp16, y = var_3285_cast_fp16)[name = tensor<string, []>("q_75_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_73_cast_fp16 = transpose(perm = k_73_perm_0, x = var_3270_cast_fp16)[name = tensor<string, []>("transpose_38")];
            tensor<fp16, [1, 4, 1, 128]> var_3287_cast_fp16 = mul(x = k_73_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3287_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = tensor<string, []>("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = tensor<string, []>("x1_75_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = tensor<string, []>("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_73_cast_fp16)[name = tensor<string, []>("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = tensor<string, []>("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = tensor<string, []>("x2_75_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = tensor<string, []>("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_73_cast_fp16)[name = tensor<string, []>("x2_75_cast_fp16")];
            tensor<fp16, []> const_38_promoted_to_fp16 = const()[name = tensor<string, []>("const_38_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_3290_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_38_promoted_to_fp16)[name = tensor<string, []>("op_3290_cast_fp16")];
            tensor<bool, []> var_3292_interleave_0 = const()[name = tensor<string, []>("op_3292_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_3292_cast_fp16 = concat(axis = var_3213, interleave = var_3292_interleave_0, values = (var_3290_cast_fp16, x1_75_cast_fp16))[name = tensor<string, []>("op_3292_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_3293_cast_fp16 = mul(x = var_3292_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3293_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_75_cast_fp16 = add(x = var_3287_cast_fp16, y = var_3293_cast_fp16)[name = tensor<string, []>("k_75_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3296_cast_fp16 = mul(x = k_cache_37_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3296_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3297_cast_fp16 = mul(x = k_75_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3297_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_37_cast_fp16 = add(x = var_3296_cast_fp16, y = var_3297_cast_fp16)[name = tensor<string, []>("k_full_37_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3300_cast_fp16 = mul(x = v_cache_37_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3300_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_3275_cast_fp16)[name = tensor<string, []>("transpose_37")];
            tensor<fp16, [1, 4, 2048, 128]> var_3301_cast_fp16 = mul(x = v_37_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3301_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_37_cast_fp16 = add(x = var_3300_cast_fp16, y = var_3301_cast_fp16)[name = tensor<string, []>("v_full_37_cast_fp16")];
            tensor<int32, [1]> var_3303_axes_0 = const()[name = tensor<string, []>("op_3303_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3303_cast_fp16 = expand_dims(axes = var_3303_axes_0, x = k_full_37_cast_fp16)[name = tensor<string, []>("op_3303_cast_fp16")];
            tensor<int32, [5]> var_3305_reps_0 = const()[name = tensor<string, []>("op_3305_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3305_cast_fp16 = tile(reps = var_3305_reps_0, x = var_3303_cast_fp16)[name = tensor<string, []>("op_3305_cast_fp16")];
            tensor<int32, [4]> var_3306 = const()[name = tensor<string, []>("op_3306"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_37_cast_fp16 = reshape(shape = var_3306, x = var_3305_cast_fp16)[name = tensor<string, []>("k_rep_37_cast_fp16")];
            tensor<int32, [1]> var_3308_axes_0 = const()[name = tensor<string, []>("op_3308_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3308_cast_fp16 = expand_dims(axes = var_3308_axes_0, x = v_full_37_cast_fp16)[name = tensor<string, []>("op_3308_cast_fp16")];
            tensor<int32, [5]> var_3310_reps_0 = const()[name = tensor<string, []>("op_3310_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3310_cast_fp16 = tile(reps = var_3310_reps_0, x = var_3308_cast_fp16)[name = tensor<string, []>("op_3310_cast_fp16")];
            tensor<int32, [4]> var_3311 = const()[name = tensor<string, []>("op_3311"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_37_cast_fp16 = reshape(shape = var_3311, x = var_3310_cast_fp16)[name = tensor<string, []>("v_rep_37_cast_fp16")];
            tensor<bool, []> var_3314_transpose_x_1 = const()[name = tensor<string, []>("op_3314_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3314_transpose_y_1 = const()[name = tensor<string, []>("op_3314_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_3314_cast_fp16 = matmul(transpose_x = var_3314_transpose_x_1, transpose_y = var_3314_transpose_y_1, x = q_75_cast_fp16, y = k_rep_37_cast_fp16)[name = tensor<string, []>("op_3314_cast_fp16")];
            tensor<fp16, []> var_3315_to_fp16 = const()[name = tensor<string, []>("op_3315_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_73_cast_fp16 = mul(x = var_3314_cast_fp16, y = var_3315_to_fp16)[name = tensor<string, []>("attn_73_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_75_cast_fp16 = add(x = attn_73_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_75_cast_fp16")];
            tensor<string, []> input_75_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_75_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_75_cast_fp16_to_fp32 = cast(dtype = input_75_cast_fp16_to_fp32_dtype_0, x = input_75_cast_fp16)[name = tensor<string, []>("cast_350")];
            tensor<fp32, [1, 12, 1, 2048]> attn_75 = softmax(axis = var_3213, x = input_75_cast_fp16_to_fp32)[name = tensor<string, []>("attn_75")];
            tensor<bool, []> out_37_transpose_x_0 = const()[name = tensor<string, []>("out_37_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_37_transpose_y_0 = const()[name = tensor<string, []>("out_37_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_75_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_75_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_75_to_fp16 = cast(dtype = attn_75_to_fp16_dtype_0, x = attn_75)[name = tensor<string, []>("cast_349")];
            tensor<fp16, [1, 12, 1, 128]> out_37_cast_fp16 = matmul(transpose_x = out_37_transpose_x_0, transpose_y = out_37_transpose_y_0, x = attn_75_to_fp16, y = v_rep_37_cast_fp16)[name = tensor<string, []>("out_37_cast_fp16")];
            tensor<int32, [4]> var_3320_perm_0 = const()[name = tensor<string, []>("op_3320_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3322 = const()[name = tensor<string, []>("op_3322"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_3320_cast_fp16 = transpose(perm = var_3320_perm_0, x = out_37_cast_fp16)[name = tensor<string, []>("transpose_36")];
            tensor<fp16, [1, 1, 1536]> x_747_cast_fp16 = reshape(shape = var_3322, x = var_3320_cast_fp16)[name = tensor<string, []>("x_747_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(385572672)))];
            tensor<fp16, [1, 1, 512]> linear_129_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = x_747_cast_fp16)[name = tensor<string, []>("linear_129_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_749_cast_fp16 = add(x = x_721_cast_fp16, y = linear_129_cast_fp16)[name = tensor<string, []>("x_749_cast_fp16")];
            tensor<string, []> x_749_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_749_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3212_promoted_3 = const()[name = tensor<string, []>("op_3212_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_749_cast_fp16_to_fp32 = cast(dtype = x_749_cast_fp16_to_fp32_dtype_0, x = x_749_cast_fp16)[name = tensor<string, []>("cast_348")];
            tensor<fp32, [1, 1, 512]> var_3333 = pow(x = x_749_cast_fp16_to_fp32, y = var_3212_promoted_3)[name = tensor<string, []>("op_3333")];
            tensor<int32, [1]> var_151_axes_0 = const()[name = tensor<string, []>("var_151_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_151_keep_dims_0 = const()[name = tensor<string, []>("var_151_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_151 = reduce_mean(axes = var_151_axes_0, keep_dims = var_151_keep_dims_0, x = var_3333)[name = tensor<string, []>("var_151")];
            tensor<string, []> var_151_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_151_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3337_to_fp16 = const()[name = tensor<string, []>("op_3337_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_151_to_fp16 = cast(dtype = var_151_to_fp16_dtype_0, x = var_151)[name = tensor<string, []>("cast_347")];
            tensor<fp16, [1, 1, 1]> var_3338_cast_fp16 = add(x = var_151_to_fp16, y = var_3337_to_fp16)[name = tensor<string, []>("op_3338_cast_fp16")];
            tensor<string, []> var_3338_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3338_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3339_epsilon_0 = const()[name = tensor<string, []>("op_3339_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3338_cast_fp16_to_fp32 = cast(dtype = var_3338_cast_fp16_to_fp32_dtype_0, x = var_3338_cast_fp16)[name = tensor<string, []>("cast_346")];
            tensor<fp32, [1, 1, 1]> var_3339 = rsqrt(epsilon = var_3339_epsilon_0, x = var_3338_cast_fp16_to_fp32)[name = tensor<string, []>("op_3339")];
            tensor<string, []> var_3339_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3339_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3339_to_fp16 = cast(dtype = var_3339_to_fp16_dtype_0, x = var_3339)[name = tensor<string, []>("cast_345")];
            tensor<fp16, [1, 1, 512]> x_755_cast_fp16 = mul(x = x_749_cast_fp16, y = var_3339_to_fp16)[name = tensor<string, []>("x_755_cast_fp16")];
            tensor<fp16, [512]> layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387145600)))];
            tensor<fp16, [1, 1, 512]> x_757_cast_fp16 = mul(x = layers_18_post_attention_layernorm_weight_to_fp16, y = x_755_cast_fp16)[name = tensor<string, []>("x_757_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_18_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(387146688)))];
            tensor<fp16, [1, 1, 1536]> linear_130_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_18_mlp_gate_proj_weight_to_fp16, x = x_757_cast_fp16)[name = tensor<string, []>("linear_130_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_3350_cast_fp16 = silu(x = linear_130_cast_fp16)[name = tensor<string, []>("op_3350_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_18_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388719616)))];
            tensor<fp16, [1, 1, 1536]> linear_131_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_18_mlp_up_proj_weight_to_fp16, x = x_757_cast_fp16)[name = tensor<string, []>("linear_131_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_759_cast_fp16 = mul(x = var_3350_cast_fp16, y = linear_131_cast_fp16)[name = tensor<string, []>("x_759_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_18_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_18_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(390292544)))];
            tensor<fp16, [1, 1, 512]> linear_132_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_18_mlp_down_proj_weight_to_fp16, x = x_759_cast_fp16)[name = tensor<string, []>("linear_132_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_761_cast_fp16 = add(x = x_749_cast_fp16, y = linear_132_cast_fp16)[name = tensor<string, []>("x_761_cast_fp16")];
            tensor<string, []> x_761_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_761_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_39_begin_0 = const()[name = tensor<string, []>("k_cache_39_begin_0"), val = tensor<int32, [5]>([19, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_39_end_0 = const()[name = tensor<string, []>("k_cache_39_end_0"), val = tensor<int32, [5]>([20, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_39_end_mask_0 = const()[name = tensor<string, []>("k_cache_39_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_39_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_39_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_39_cast_fp16 = slice_by_index(begin = k_cache_39_begin_0, end = k_cache_39_end_0, end_mask = k_cache_39_end_mask_0, squeeze_mask = k_cache_39_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_39_cast_fp16")];
            tensor<int32, [5]> v_cache_39_begin_0 = const()[name = tensor<string, []>("v_cache_39_begin_0"), val = tensor<int32, [5]>([19, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_39_end_0 = const()[name = tensor<string, []>("v_cache_39_end_0"), val = tensor<int32, [5]>([20, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_39_end_mask_0 = const()[name = tensor<string, []>("v_cache_39_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_39_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_39_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_39_cast_fp16 = slice_by_index(begin = v_cache_39_begin_0, end = v_cache_39_end_0, end_mask = v_cache_39_end_mask_0, squeeze_mask = v_cache_39_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_39_cast_fp16")];
            tensor<int32, []> var_3382 = const()[name = tensor<string, []>("op_3382"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_3381_promoted = const()[name = tensor<string, []>("op_3381_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_761_cast_fp16_to_fp32 = cast(dtype = x_761_cast_fp16_to_fp32_dtype_0, x = x_761_cast_fp16)[name = tensor<string, []>("cast_344")];
            tensor<fp32, [1, 1, 512]> var_3391 = pow(x = x_761_cast_fp16_to_fp32, y = var_3381_promoted)[name = tensor<string, []>("op_3391")];
            tensor<int32, [1]> var_153_axes_0 = const()[name = tensor<string, []>("var_153_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_153_keep_dims_0 = const()[name = tensor<string, []>("var_153_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_153 = reduce_mean(axes = var_153_axes_0, keep_dims = var_153_keep_dims_0, x = var_3391)[name = tensor<string, []>("var_153")];
            tensor<string, []> var_153_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_153_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3395_to_fp16 = const()[name = tensor<string, []>("op_3395_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_153_to_fp16 = cast(dtype = var_153_to_fp16_dtype_0, x = var_153)[name = tensor<string, []>("cast_343")];
            tensor<fp16, [1, 1, 1]> var_3396_cast_fp16 = add(x = var_153_to_fp16, y = var_3395_to_fp16)[name = tensor<string, []>("op_3396_cast_fp16")];
            tensor<string, []> var_3396_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3396_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3397_epsilon_0 = const()[name = tensor<string, []>("op_3397_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3396_cast_fp16_to_fp32 = cast(dtype = var_3396_cast_fp16_to_fp32_dtype_0, x = var_3396_cast_fp16)[name = tensor<string, []>("cast_342")];
            tensor<fp32, [1, 1, 1]> var_3397 = rsqrt(epsilon = var_3397_epsilon_0, x = var_3396_cast_fp16_to_fp32)[name = tensor<string, []>("op_3397")];
            tensor<string, []> var_3397_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3397_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3397_to_fp16 = cast(dtype = var_3397_to_fp16_dtype_0, x = var_3397)[name = tensor<string, []>("cast_341")];
            tensor<fp16, [1, 1, 512]> x_767_cast_fp16 = mul(x = x_761_cast_fp16, y = var_3397_to_fp16)[name = tensor<string, []>("x_767_cast_fp16")];
            tensor<fp16, [512]> layers_19_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391865472)))];
            tensor<fp16, [1, 1, 512]> x_769_cast_fp16 = mul(x = layers_19_input_layernorm_weight_to_fp16, y = x_767_cast_fp16)[name = tensor<string, []>("x_769_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(391866560)))];
            tensor<fp16, [1, 1, 1536]> linear_133_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = x_769_cast_fp16)[name = tensor<string, []>("linear_133_cast_fp16")];
            tensor<int32, [4]> var_3411 = const()[name = tensor<string, []>("op_3411"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_771_cast_fp16 = reshape(shape = var_3411, x = linear_133_cast_fp16)[name = tensor<string, []>("x_771_cast_fp16")];
            tensor<string, []> x_771_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_771_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3381_promoted_1 = const()[name = tensor<string, []>("op_3381_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_771_cast_fp16_to_fp32 = cast(dtype = x_771_cast_fp16_to_fp32_dtype_0, x = x_771_cast_fp16)[name = tensor<string, []>("cast_340")];
            tensor<fp32, [1, 1, 12, 128]> var_3415 = pow(x = x_771_cast_fp16_to_fp32, y = var_3381_promoted_1)[name = tensor<string, []>("op_3415")];
            tensor<int32, [1]> var_155_axes_0 = const()[name = tensor<string, []>("var_155_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_155_keep_dims_0 = const()[name = tensor<string, []>("var_155_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_155 = reduce_mean(axes = var_155_axes_0, keep_dims = var_155_keep_dims_0, x = var_3415)[name = tensor<string, []>("var_155")];
            tensor<string, []> var_155_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_155_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3419_to_fp16 = const()[name = tensor<string, []>("op_3419_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_155_to_fp16 = cast(dtype = var_155_to_fp16_dtype_0, x = var_155)[name = tensor<string, []>("cast_339")];
            tensor<fp16, [1, 1, 12, 1]> var_3420_cast_fp16 = add(x = var_155_to_fp16, y = var_3419_to_fp16)[name = tensor<string, []>("op_3420_cast_fp16")];
            tensor<string, []> var_3420_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3420_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3421_epsilon_0 = const()[name = tensor<string, []>("op_3421_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_3420_cast_fp16_to_fp32 = cast(dtype = var_3420_cast_fp16_to_fp32_dtype_0, x = var_3420_cast_fp16)[name = tensor<string, []>("cast_338")];
            tensor<fp32, [1, 1, 12, 1]> var_3421 = rsqrt(epsilon = var_3421_epsilon_0, x = var_3420_cast_fp16_to_fp32)[name = tensor<string, []>("op_3421")];
            tensor<string, []> var_3421_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3421_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_3421_to_fp16 = cast(dtype = var_3421_to_fp16_dtype_0, x = var_3421)[name = tensor<string, []>("cast_337")];
            tensor<fp16, [1, 1, 12, 128]> x_777_cast_fp16 = mul(x = x_771_cast_fp16, y = var_3421_to_fp16)[name = tensor<string, []>("x_777_cast_fp16")];
            tensor<fp16, [128]> layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393439488)))];
            tensor<fp16, [1, 1, 12, 128]> var_3423_cast_fp16 = mul(x = layers_19_self_attn_q_norm_weight_to_fp16, y = x_777_cast_fp16)[name = tensor<string, []>("op_3423_cast_fp16")];
            tensor<int32, [4]> q_77_perm_0 = const()[name = tensor<string, []>("q_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393439808)))];
            tensor<fp16, [1, 1, 512]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = x_769_cast_fp16)[name = tensor<string, []>("linear_134_cast_fp16")];
            tensor<int32, [4]> var_3427 = const()[name = tensor<string, []>("op_3427"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_779_cast_fp16 = reshape(shape = var_3427, x = linear_134_cast_fp16)[name = tensor<string, []>("x_779_cast_fp16")];
            tensor<string, []> x_779_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_779_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3381_promoted_2 = const()[name = tensor<string, []>("op_3381_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_779_cast_fp16_to_fp32 = cast(dtype = x_779_cast_fp16_to_fp32_dtype_0, x = x_779_cast_fp16)[name = tensor<string, []>("cast_336")];
            tensor<fp32, [1, 1, 4, 128]> var_3431 = pow(x = x_779_cast_fp16_to_fp32, y = var_3381_promoted_2)[name = tensor<string, []>("op_3431")];
            tensor<int32, [1]> var_157_axes_0 = const()[name = tensor<string, []>("var_157_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_157_keep_dims_0 = const()[name = tensor<string, []>("var_157_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_157 = reduce_mean(axes = var_157_axes_0, keep_dims = var_157_keep_dims_0, x = var_3431)[name = tensor<string, []>("var_157")];
            tensor<string, []> var_157_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_157_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3435_to_fp16 = const()[name = tensor<string, []>("op_3435_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_157_to_fp16 = cast(dtype = var_157_to_fp16_dtype_0, x = var_157)[name = tensor<string, []>("cast_335")];
            tensor<fp16, [1, 1, 4, 1]> var_3436_cast_fp16 = add(x = var_157_to_fp16, y = var_3435_to_fp16)[name = tensor<string, []>("op_3436_cast_fp16")];
            tensor<string, []> var_3436_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3436_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3437_epsilon_0 = const()[name = tensor<string, []>("op_3437_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_3436_cast_fp16_to_fp32 = cast(dtype = var_3436_cast_fp16_to_fp32_dtype_0, x = var_3436_cast_fp16)[name = tensor<string, []>("cast_334")];
            tensor<fp32, [1, 1, 4, 1]> var_3437 = rsqrt(epsilon = var_3437_epsilon_0, x = var_3436_cast_fp16_to_fp32)[name = tensor<string, []>("op_3437")];
            tensor<string, []> var_3437_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3437_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_3437_to_fp16 = cast(dtype = var_3437_to_fp16_dtype_0, x = var_3437)[name = tensor<string, []>("cast_333")];
            tensor<fp16, [1, 1, 4, 128]> x_785_cast_fp16 = mul(x = x_779_cast_fp16, y = var_3437_to_fp16)[name = tensor<string, []>("x_785_cast_fp16")];
            tensor<fp16, [128]> layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393964160)))];
            tensor<fp16, [1, 1, 4, 128]> var_3439_cast_fp16 = mul(x = layers_19_self_attn_k_norm_weight_to_fp16, y = x_785_cast_fp16)[name = tensor<string, []>("op_3439_cast_fp16")];
            tensor<int32, [4]> k_77_perm_0 = const()[name = tensor<string, []>("k_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393964480)))];
            tensor<fp16, [1, 1, 512]> linear_135_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = x_769_cast_fp16)[name = tensor<string, []>("linear_135_cast_fp16")];
            tensor<int32, [4]> var_3443 = const()[name = tensor<string, []>("op_3443"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_3444_cast_fp16 = reshape(shape = var_3443, x = linear_135_cast_fp16)[name = tensor<string, []>("op_3444_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = tensor<string, []>("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_77_cast_fp16 = transpose(perm = q_77_perm_0, x = var_3423_cast_fp16)[name = tensor<string, []>("transpose_35")];
            tensor<fp16, [1, 12, 1, 128]> var_3448_cast_fp16 = mul(x = q_77_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3448_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = tensor<string, []>("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = tensor<string, []>("x1_77_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = tensor<string, []>("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_77_cast_fp16)[name = tensor<string, []>("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = tensor<string, []>("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = tensor<string, []>("x2_77_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = tensor<string, []>("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_77_cast_fp16)[name = tensor<string, []>("x2_77_cast_fp16")];
            tensor<fp16, []> const_39_promoted_to_fp16 = const()[name = tensor<string, []>("const_39_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_3451_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_39_promoted_to_fp16)[name = tensor<string, []>("op_3451_cast_fp16")];
            tensor<bool, []> var_3453_interleave_0 = const()[name = tensor<string, []>("op_3453_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_3453_cast_fp16 = concat(axis = var_3382, interleave = var_3453_interleave_0, values = (var_3451_cast_fp16, x1_77_cast_fp16))[name = tensor<string, []>("op_3453_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_3454_cast_fp16 = mul(x = var_3453_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3454_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_79_cast_fp16 = add(x = var_3448_cast_fp16, y = var_3454_cast_fp16)[name = tensor<string, []>("q_79_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_77_cast_fp16 = transpose(perm = k_77_perm_0, x = var_3439_cast_fp16)[name = tensor<string, []>("transpose_34")];
            tensor<fp16, [1, 4, 1, 128]> var_3456_cast_fp16 = mul(x = k_77_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3456_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = tensor<string, []>("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = tensor<string, []>("x1_79_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = tensor<string, []>("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_77_cast_fp16)[name = tensor<string, []>("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = tensor<string, []>("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = tensor<string, []>("x2_79_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = tensor<string, []>("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_77_cast_fp16)[name = tensor<string, []>("x2_79_cast_fp16")];
            tensor<fp16, []> const_40_promoted_to_fp16 = const()[name = tensor<string, []>("const_40_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_3459_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_40_promoted_to_fp16)[name = tensor<string, []>("op_3459_cast_fp16")];
            tensor<bool, []> var_3461_interleave_0 = const()[name = tensor<string, []>("op_3461_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_3461_cast_fp16 = concat(axis = var_3382, interleave = var_3461_interleave_0, values = (var_3459_cast_fp16, x1_79_cast_fp16))[name = tensor<string, []>("op_3461_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_3462_cast_fp16 = mul(x = var_3461_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3462_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_79_cast_fp16 = add(x = var_3456_cast_fp16, y = var_3462_cast_fp16)[name = tensor<string, []>("k_79_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3465_cast_fp16 = mul(x = k_cache_39_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3465_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3466_cast_fp16 = mul(x = k_79_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3466_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_39_cast_fp16 = add(x = var_3465_cast_fp16, y = var_3466_cast_fp16)[name = tensor<string, []>("k_full_39_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3469_cast_fp16 = mul(x = v_cache_39_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3469_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_3444_cast_fp16)[name = tensor<string, []>("transpose_33")];
            tensor<fp16, [1, 4, 2048, 128]> var_3470_cast_fp16 = mul(x = v_39_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3470_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_39_cast_fp16 = add(x = var_3469_cast_fp16, y = var_3470_cast_fp16)[name = tensor<string, []>("v_full_39_cast_fp16")];
            tensor<int32, [1]> var_3472_axes_0 = const()[name = tensor<string, []>("op_3472_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3472_cast_fp16 = expand_dims(axes = var_3472_axes_0, x = k_full_39_cast_fp16)[name = tensor<string, []>("op_3472_cast_fp16")];
            tensor<int32, [5]> var_3474_reps_0 = const()[name = tensor<string, []>("op_3474_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3474_cast_fp16 = tile(reps = var_3474_reps_0, x = var_3472_cast_fp16)[name = tensor<string, []>("op_3474_cast_fp16")];
            tensor<int32, [4]> var_3475 = const()[name = tensor<string, []>("op_3475"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_39_cast_fp16 = reshape(shape = var_3475, x = var_3474_cast_fp16)[name = tensor<string, []>("k_rep_39_cast_fp16")];
            tensor<int32, [1]> var_3477_axes_0 = const()[name = tensor<string, []>("op_3477_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3477_cast_fp16 = expand_dims(axes = var_3477_axes_0, x = v_full_39_cast_fp16)[name = tensor<string, []>("op_3477_cast_fp16")];
            tensor<int32, [5]> var_3479_reps_0 = const()[name = tensor<string, []>("op_3479_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3479_cast_fp16 = tile(reps = var_3479_reps_0, x = var_3477_cast_fp16)[name = tensor<string, []>("op_3479_cast_fp16")];
            tensor<int32, [4]> var_3480 = const()[name = tensor<string, []>("op_3480"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_39_cast_fp16 = reshape(shape = var_3480, x = var_3479_cast_fp16)[name = tensor<string, []>("v_rep_39_cast_fp16")];
            tensor<bool, []> var_3483_transpose_x_1 = const()[name = tensor<string, []>("op_3483_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3483_transpose_y_1 = const()[name = tensor<string, []>("op_3483_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_3483_cast_fp16 = matmul(transpose_x = var_3483_transpose_x_1, transpose_y = var_3483_transpose_y_1, x = q_79_cast_fp16, y = k_rep_39_cast_fp16)[name = tensor<string, []>("op_3483_cast_fp16")];
            tensor<fp16, []> var_3484_to_fp16 = const()[name = tensor<string, []>("op_3484_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_77_cast_fp16 = mul(x = var_3483_cast_fp16, y = var_3484_to_fp16)[name = tensor<string, []>("attn_77_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_79_cast_fp16 = add(x = attn_77_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_79_cast_fp16")];
            tensor<string, []> input_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_79_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_79_cast_fp16_to_fp32 = cast(dtype = input_79_cast_fp16_to_fp32_dtype_0, x = input_79_cast_fp16)[name = tensor<string, []>("cast_332")];
            tensor<fp32, [1, 12, 1, 2048]> attn_79 = softmax(axis = var_3382, x = input_79_cast_fp16_to_fp32)[name = tensor<string, []>("attn_79")];
            tensor<bool, []> out_39_transpose_x_0 = const()[name = tensor<string, []>("out_39_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_39_transpose_y_0 = const()[name = tensor<string, []>("out_39_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_79_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_79_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_79_to_fp16 = cast(dtype = attn_79_to_fp16_dtype_0, x = attn_79)[name = tensor<string, []>("cast_331")];
            tensor<fp16, [1, 12, 1, 128]> out_39_cast_fp16 = matmul(transpose_x = out_39_transpose_x_0, transpose_y = out_39_transpose_y_0, x = attn_79_to_fp16, y = v_rep_39_cast_fp16)[name = tensor<string, []>("out_39_cast_fp16")];
            tensor<int32, [4]> var_3489_perm_0 = const()[name = tensor<string, []>("op_3489_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3491 = const()[name = tensor<string, []>("op_3491"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_3489_cast_fp16 = transpose(perm = var_3489_perm_0, x = out_39_cast_fp16)[name = tensor<string, []>("transpose_32")];
            tensor<fp16, [1, 1, 1536]> x_787_cast_fp16 = reshape(shape = var_3491, x = var_3489_cast_fp16)[name = tensor<string, []>("x_787_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(394488832)))];
            tensor<fp16, [1, 1, 512]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = x_787_cast_fp16)[name = tensor<string, []>("linear_136_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_789_cast_fp16 = add(x = x_761_cast_fp16, y = linear_136_cast_fp16)[name = tensor<string, []>("x_789_cast_fp16")];
            tensor<string, []> x_789_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_789_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3381_promoted_3 = const()[name = tensor<string, []>("op_3381_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_789_cast_fp16_to_fp32 = cast(dtype = x_789_cast_fp16_to_fp32_dtype_0, x = x_789_cast_fp16)[name = tensor<string, []>("cast_330")];
            tensor<fp32, [1, 1, 512]> var_3502 = pow(x = x_789_cast_fp16_to_fp32, y = var_3381_promoted_3)[name = tensor<string, []>("op_3502")];
            tensor<int32, [1]> var_159_axes_0 = const()[name = tensor<string, []>("var_159_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_159_keep_dims_0 = const()[name = tensor<string, []>("var_159_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_159 = reduce_mean(axes = var_159_axes_0, keep_dims = var_159_keep_dims_0, x = var_3502)[name = tensor<string, []>("var_159")];
            tensor<string, []> var_159_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_159_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3506_to_fp16 = const()[name = tensor<string, []>("op_3506_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_159_to_fp16 = cast(dtype = var_159_to_fp16_dtype_0, x = var_159)[name = tensor<string, []>("cast_329")];
            tensor<fp16, [1, 1, 1]> var_3507_cast_fp16 = add(x = var_159_to_fp16, y = var_3506_to_fp16)[name = tensor<string, []>("op_3507_cast_fp16")];
            tensor<string, []> var_3507_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3507_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3508_epsilon_0 = const()[name = tensor<string, []>("op_3508_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3507_cast_fp16_to_fp32 = cast(dtype = var_3507_cast_fp16_to_fp32_dtype_0, x = var_3507_cast_fp16)[name = tensor<string, []>("cast_328")];
            tensor<fp32, [1, 1, 1]> var_3508 = rsqrt(epsilon = var_3508_epsilon_0, x = var_3507_cast_fp16_to_fp32)[name = tensor<string, []>("op_3508")];
            tensor<string, []> var_3508_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3508_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3508_to_fp16 = cast(dtype = var_3508_to_fp16_dtype_0, x = var_3508)[name = tensor<string, []>("cast_327")];
            tensor<fp16, [1, 1, 512]> x_795_cast_fp16 = mul(x = x_789_cast_fp16, y = var_3508_to_fp16)[name = tensor<string, []>("x_795_cast_fp16")];
            tensor<fp16, [512]> layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(396061760)))];
            tensor<fp16, [1, 1, 512]> x_797_cast_fp16 = mul(x = layers_19_post_attention_layernorm_weight_to_fp16, y = x_795_cast_fp16)[name = tensor<string, []>("x_797_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_19_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(396062848)))];
            tensor<fp16, [1, 1, 1536]> linear_137_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_19_mlp_gate_proj_weight_to_fp16, x = x_797_cast_fp16)[name = tensor<string, []>("linear_137_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_3519_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor<string, []>("op_3519_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_19_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397635776)))];
            tensor<fp16, [1, 1, 1536]> linear_138_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_19_mlp_up_proj_weight_to_fp16, x = x_797_cast_fp16)[name = tensor<string, []>("linear_138_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_799_cast_fp16 = mul(x = var_3519_cast_fp16, y = linear_138_cast_fp16)[name = tensor<string, []>("x_799_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_19_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_19_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(399208704)))];
            tensor<fp16, [1, 1, 512]> linear_139_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_19_mlp_down_proj_weight_to_fp16, x = x_799_cast_fp16)[name = tensor<string, []>("linear_139_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_801_cast_fp16 = add(x = x_789_cast_fp16, y = linear_139_cast_fp16)[name = tensor<string, []>("x_801_cast_fp16")];
            tensor<string, []> x_801_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_801_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_41_begin_0 = const()[name = tensor<string, []>("k_cache_41_begin_0"), val = tensor<int32, [5]>([20, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_41_end_0 = const()[name = tensor<string, []>("k_cache_41_end_0"), val = tensor<int32, [5]>([21, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_41_end_mask_0 = const()[name = tensor<string, []>("k_cache_41_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_41_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_41_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_41_cast_fp16 = slice_by_index(begin = k_cache_41_begin_0, end = k_cache_41_end_0, end_mask = k_cache_41_end_mask_0, squeeze_mask = k_cache_41_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_41_cast_fp16")];
            tensor<int32, [5]> v_cache_41_begin_0 = const()[name = tensor<string, []>("v_cache_41_begin_0"), val = tensor<int32, [5]>([20, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_41_end_0 = const()[name = tensor<string, []>("v_cache_41_end_0"), val = tensor<int32, [5]>([21, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_41_end_mask_0 = const()[name = tensor<string, []>("v_cache_41_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_41_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_41_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_41_cast_fp16 = slice_by_index(begin = v_cache_41_begin_0, end = v_cache_41_end_0, end_mask = v_cache_41_end_mask_0, squeeze_mask = v_cache_41_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_41_cast_fp16")];
            tensor<int32, []> var_3551 = const()[name = tensor<string, []>("op_3551"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_3550_promoted = const()[name = tensor<string, []>("op_3550_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_801_cast_fp16_to_fp32 = cast(dtype = x_801_cast_fp16_to_fp32_dtype_0, x = x_801_cast_fp16)[name = tensor<string, []>("cast_326")];
            tensor<fp32, [1, 1, 512]> var_3560 = pow(x = x_801_cast_fp16_to_fp32, y = var_3550_promoted)[name = tensor<string, []>("op_3560")];
            tensor<int32, [1]> var_161_axes_0 = const()[name = tensor<string, []>("var_161_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_161_keep_dims_0 = const()[name = tensor<string, []>("var_161_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_161 = reduce_mean(axes = var_161_axes_0, keep_dims = var_161_keep_dims_0, x = var_3560)[name = tensor<string, []>("var_161")];
            tensor<string, []> var_161_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_161_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3564_to_fp16 = const()[name = tensor<string, []>("op_3564_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_161_to_fp16 = cast(dtype = var_161_to_fp16_dtype_0, x = var_161)[name = tensor<string, []>("cast_325")];
            tensor<fp16, [1, 1, 1]> var_3565_cast_fp16 = add(x = var_161_to_fp16, y = var_3564_to_fp16)[name = tensor<string, []>("op_3565_cast_fp16")];
            tensor<string, []> var_3565_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3565_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3566_epsilon_0 = const()[name = tensor<string, []>("op_3566_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3565_cast_fp16_to_fp32 = cast(dtype = var_3565_cast_fp16_to_fp32_dtype_0, x = var_3565_cast_fp16)[name = tensor<string, []>("cast_324")];
            tensor<fp32, [1, 1, 1]> var_3566 = rsqrt(epsilon = var_3566_epsilon_0, x = var_3565_cast_fp16_to_fp32)[name = tensor<string, []>("op_3566")];
            tensor<string, []> var_3566_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3566_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3566_to_fp16 = cast(dtype = var_3566_to_fp16_dtype_0, x = var_3566)[name = tensor<string, []>("cast_323")];
            tensor<fp16, [1, 1, 512]> x_807_cast_fp16 = mul(x = x_801_cast_fp16, y = var_3566_to_fp16)[name = tensor<string, []>("x_807_cast_fp16")];
            tensor<fp16, [512]> layers_20_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(400781632)))];
            tensor<fp16, [1, 1, 512]> x_809_cast_fp16 = mul(x = layers_20_input_layernorm_weight_to_fp16, y = x_807_cast_fp16)[name = tensor<string, []>("x_809_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(400782720)))];
            tensor<fp16, [1, 1, 1536]> linear_140_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = x_809_cast_fp16)[name = tensor<string, []>("linear_140_cast_fp16")];
            tensor<int32, [4]> var_3580 = const()[name = tensor<string, []>("op_3580"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_811_cast_fp16 = reshape(shape = var_3580, x = linear_140_cast_fp16)[name = tensor<string, []>("x_811_cast_fp16")];
            tensor<string, []> x_811_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_811_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3550_promoted_1 = const()[name = tensor<string, []>("op_3550_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_811_cast_fp16_to_fp32 = cast(dtype = x_811_cast_fp16_to_fp32_dtype_0, x = x_811_cast_fp16)[name = tensor<string, []>("cast_322")];
            tensor<fp32, [1, 1, 12, 128]> var_3584 = pow(x = x_811_cast_fp16_to_fp32, y = var_3550_promoted_1)[name = tensor<string, []>("op_3584")];
            tensor<int32, [1]> var_163_axes_0 = const()[name = tensor<string, []>("var_163_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_163_keep_dims_0 = const()[name = tensor<string, []>("var_163_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_163 = reduce_mean(axes = var_163_axes_0, keep_dims = var_163_keep_dims_0, x = var_3584)[name = tensor<string, []>("var_163")];
            tensor<string, []> var_163_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_163_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3588_to_fp16 = const()[name = tensor<string, []>("op_3588_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_163_to_fp16_0 = cast(dtype = var_163_to_fp16_dtype_0, x = var_163)[name = tensor<string, []>("cast_321")];
            tensor<fp16, [1, 1, 12, 1]> var_3589_cast_fp16 = add(x = var_163_to_fp16_0, y = var_3588_to_fp16)[name = tensor<string, []>("op_3589_cast_fp16")];
            tensor<string, []> var_3589_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3589_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3590_epsilon_0 = const()[name = tensor<string, []>("op_3590_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_3589_cast_fp16_to_fp32 = cast(dtype = var_3589_cast_fp16_to_fp32_dtype_0, x = var_3589_cast_fp16)[name = tensor<string, []>("cast_320")];
            tensor<fp32, [1, 1, 12, 1]> var_3590 = rsqrt(epsilon = var_3590_epsilon_0, x = var_3589_cast_fp16_to_fp32)[name = tensor<string, []>("op_3590")];
            tensor<string, []> var_3590_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3590_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_3590_to_fp16 = cast(dtype = var_3590_to_fp16_dtype_0, x = var_3590)[name = tensor<string, []>("cast_319")];
            tensor<fp16, [1, 1, 12, 128]> x_817_cast_fp16 = mul(x = x_811_cast_fp16, y = var_3590_to_fp16)[name = tensor<string, []>("x_817_cast_fp16")];
            tensor<fp16, [128]> layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402355648)))];
            tensor<fp16, [1, 1, 12, 128]> var_3592_cast_fp16 = mul(x = layers_20_self_attn_q_norm_weight_to_fp16, y = x_817_cast_fp16)[name = tensor<string, []>("op_3592_cast_fp16")];
            tensor<int32, [4]> q_81_perm_0 = const()[name = tensor<string, []>("q_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402355968)))];
            tensor<fp16, [1, 1, 512]> linear_141_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = x_809_cast_fp16)[name = tensor<string, []>("linear_141_cast_fp16")];
            tensor<int32, [4]> var_3596 = const()[name = tensor<string, []>("op_3596"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_819_cast_fp16 = reshape(shape = var_3596, x = linear_141_cast_fp16)[name = tensor<string, []>("x_819_cast_fp16")];
            tensor<string, []> x_819_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_819_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3550_promoted_2 = const()[name = tensor<string, []>("op_3550_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_819_cast_fp16_to_fp32 = cast(dtype = x_819_cast_fp16_to_fp32_dtype_0, x = x_819_cast_fp16)[name = tensor<string, []>("cast_318")];
            tensor<fp32, [1, 1, 4, 128]> var_3600 = pow(x = x_819_cast_fp16_to_fp32, y = var_3550_promoted_2)[name = tensor<string, []>("op_3600")];
            tensor<int32, [1]> var_165_axes_0 = const()[name = tensor<string, []>("var_165_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_165_keep_dims_0 = const()[name = tensor<string, []>("var_165_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_165 = reduce_mean(axes = var_165_axes_0, keep_dims = var_165_keep_dims_0, x = var_3600)[name = tensor<string, []>("var_165")];
            tensor<string, []> var_165_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_165_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3604_to_fp16 = const()[name = tensor<string, []>("op_3604_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_165_to_fp16 = cast(dtype = var_165_to_fp16_dtype_0, x = var_165)[name = tensor<string, []>("cast_317")];
            tensor<fp16, [1, 1, 4, 1]> var_3605_cast_fp16 = add(x = var_165_to_fp16, y = var_3604_to_fp16)[name = tensor<string, []>("op_3605_cast_fp16")];
            tensor<string, []> var_3605_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3605_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3606_epsilon_0 = const()[name = tensor<string, []>("op_3606_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_3605_cast_fp16_to_fp32 = cast(dtype = var_3605_cast_fp16_to_fp32_dtype_0, x = var_3605_cast_fp16)[name = tensor<string, []>("cast_316")];
            tensor<fp32, [1, 1, 4, 1]> var_3606 = rsqrt(epsilon = var_3606_epsilon_0, x = var_3605_cast_fp16_to_fp32)[name = tensor<string, []>("op_3606")];
            tensor<string, []> var_3606_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3606_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_3606_to_fp16 = cast(dtype = var_3606_to_fp16_dtype_0, x = var_3606)[name = tensor<string, []>("cast_315")];
            tensor<fp16, [1, 1, 4, 128]> x_825_cast_fp16 = mul(x = x_819_cast_fp16, y = var_3606_to_fp16)[name = tensor<string, []>("x_825_cast_fp16")];
            tensor<fp16, [128]> layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402880320)))];
            tensor<fp16, [1, 1, 4, 128]> var_3608_cast_fp16 = mul(x = layers_20_self_attn_k_norm_weight_to_fp16, y = x_825_cast_fp16)[name = tensor<string, []>("op_3608_cast_fp16")];
            tensor<int32, [4]> k_81_perm_0 = const()[name = tensor<string, []>("k_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402880640)))];
            tensor<fp16, [1, 1, 512]> linear_142_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = x_809_cast_fp16)[name = tensor<string, []>("linear_142_cast_fp16")];
            tensor<int32, [4]> var_3612 = const()[name = tensor<string, []>("op_3612"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_3613_cast_fp16 = reshape(shape = var_3612, x = linear_142_cast_fp16)[name = tensor<string, []>("op_3613_cast_fp16")];
            tensor<int32, [4]> v_41_perm_0 = const()[name = tensor<string, []>("v_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_81_cast_fp16 = transpose(perm = q_81_perm_0, x = var_3592_cast_fp16)[name = tensor<string, []>("transpose_31")];
            tensor<fp16, [1, 12, 1, 128]> var_3617_cast_fp16 = mul(x = q_81_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3617_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = tensor<string, []>("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = tensor<string, []>("x1_81_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = tensor<string, []>("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_81_cast_fp16)[name = tensor<string, []>("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = tensor<string, []>("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = tensor<string, []>("x2_81_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = tensor<string, []>("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_81_cast_fp16)[name = tensor<string, []>("x2_81_cast_fp16")];
            tensor<fp16, []> const_41_promoted_to_fp16 = const()[name = tensor<string, []>("const_41_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_3620_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_41_promoted_to_fp16)[name = tensor<string, []>("op_3620_cast_fp16")];
            tensor<bool, []> var_3622_interleave_0 = const()[name = tensor<string, []>("op_3622_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_3622_cast_fp16 = concat(axis = var_3551, interleave = var_3622_interleave_0, values = (var_3620_cast_fp16, x1_81_cast_fp16))[name = tensor<string, []>("op_3622_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_3623_cast_fp16 = mul(x = var_3622_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3623_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_83_cast_fp16 = add(x = var_3617_cast_fp16, y = var_3623_cast_fp16)[name = tensor<string, []>("q_83_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_81_cast_fp16 = transpose(perm = k_81_perm_0, x = var_3608_cast_fp16)[name = tensor<string, []>("transpose_30")];
            tensor<fp16, [1, 4, 1, 128]> var_3625_cast_fp16 = mul(x = k_81_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3625_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = tensor<string, []>("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = tensor<string, []>("x1_83_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = tensor<string, []>("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_81_cast_fp16)[name = tensor<string, []>("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = tensor<string, []>("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = tensor<string, []>("x2_83_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = tensor<string, []>("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_81_cast_fp16)[name = tensor<string, []>("x2_83_cast_fp16")];
            tensor<fp16, []> const_42_promoted_to_fp16 = const()[name = tensor<string, []>("const_42_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_3628_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_42_promoted_to_fp16)[name = tensor<string, []>("op_3628_cast_fp16")];
            tensor<bool, []> var_3630_interleave_0 = const()[name = tensor<string, []>("op_3630_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_3630_cast_fp16 = concat(axis = var_3551, interleave = var_3630_interleave_0, values = (var_3628_cast_fp16, x1_83_cast_fp16))[name = tensor<string, []>("op_3630_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_3631_cast_fp16 = mul(x = var_3630_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3631_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_83_cast_fp16 = add(x = var_3625_cast_fp16, y = var_3631_cast_fp16)[name = tensor<string, []>("k_83_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3634_cast_fp16 = mul(x = k_cache_41_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3634_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3635_cast_fp16 = mul(x = k_83_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3635_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_41_cast_fp16 = add(x = var_3634_cast_fp16, y = var_3635_cast_fp16)[name = tensor<string, []>("k_full_41_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3638_cast_fp16 = mul(x = v_cache_41_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3638_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = var_3613_cast_fp16)[name = tensor<string, []>("transpose_29")];
            tensor<fp16, [1, 4, 2048, 128]> var_3639_cast_fp16 = mul(x = v_41_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3639_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_41_cast_fp16 = add(x = var_3638_cast_fp16, y = var_3639_cast_fp16)[name = tensor<string, []>("v_full_41_cast_fp16")];
            tensor<int32, [1]> var_3641_axes_0 = const()[name = tensor<string, []>("op_3641_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3641_cast_fp16 = expand_dims(axes = var_3641_axes_0, x = k_full_41_cast_fp16)[name = tensor<string, []>("op_3641_cast_fp16")];
            tensor<int32, [5]> var_3643_reps_0 = const()[name = tensor<string, []>("op_3643_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3643_cast_fp16 = tile(reps = var_3643_reps_0, x = var_3641_cast_fp16)[name = tensor<string, []>("op_3643_cast_fp16")];
            tensor<int32, [4]> var_3644 = const()[name = tensor<string, []>("op_3644"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_41_cast_fp16 = reshape(shape = var_3644, x = var_3643_cast_fp16)[name = tensor<string, []>("k_rep_41_cast_fp16")];
            tensor<int32, [1]> var_3646_axes_0 = const()[name = tensor<string, []>("op_3646_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3646_cast_fp16 = expand_dims(axes = var_3646_axes_0, x = v_full_41_cast_fp16)[name = tensor<string, []>("op_3646_cast_fp16")];
            tensor<int32, [5]> var_3648_reps_0 = const()[name = tensor<string, []>("op_3648_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3648_cast_fp16 = tile(reps = var_3648_reps_0, x = var_3646_cast_fp16)[name = tensor<string, []>("op_3648_cast_fp16")];
            tensor<int32, [4]> var_3649 = const()[name = tensor<string, []>("op_3649"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_41_cast_fp16 = reshape(shape = var_3649, x = var_3648_cast_fp16)[name = tensor<string, []>("v_rep_41_cast_fp16")];
            tensor<bool, []> var_3652_transpose_x_1 = const()[name = tensor<string, []>("op_3652_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3652_transpose_y_1 = const()[name = tensor<string, []>("op_3652_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_3652_cast_fp16 = matmul(transpose_x = var_3652_transpose_x_1, transpose_y = var_3652_transpose_y_1, x = q_83_cast_fp16, y = k_rep_41_cast_fp16)[name = tensor<string, []>("op_3652_cast_fp16")];
            tensor<fp16, []> var_3653_to_fp16 = const()[name = tensor<string, []>("op_3653_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_81_cast_fp16 = mul(x = var_3652_cast_fp16, y = var_3653_to_fp16)[name = tensor<string, []>("attn_81_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_83_cast_fp16 = add(x = attn_81_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_83_cast_fp16")];
            tensor<string, []> input_83_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_83_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_83_cast_fp16_to_fp32 = cast(dtype = input_83_cast_fp16_to_fp32_dtype_0, x = input_83_cast_fp16)[name = tensor<string, []>("cast_314")];
            tensor<fp32, [1, 12, 1, 2048]> attn_83 = softmax(axis = var_3551, x = input_83_cast_fp16_to_fp32)[name = tensor<string, []>("attn_83")];
            tensor<bool, []> out_41_transpose_x_0 = const()[name = tensor<string, []>("out_41_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_41_transpose_y_0 = const()[name = tensor<string, []>("out_41_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_83_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_83_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_83_to_fp16 = cast(dtype = attn_83_to_fp16_dtype_0, x = attn_83)[name = tensor<string, []>("cast_313")];
            tensor<fp16, [1, 12, 1, 128]> out_41_cast_fp16 = matmul(transpose_x = out_41_transpose_x_0, transpose_y = out_41_transpose_y_0, x = attn_83_to_fp16, y = v_rep_41_cast_fp16)[name = tensor<string, []>("out_41_cast_fp16")];
            tensor<int32, [4]> var_3658_perm_0 = const()[name = tensor<string, []>("op_3658_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3660 = const()[name = tensor<string, []>("op_3660"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_3658_cast_fp16 = transpose(perm = var_3658_perm_0, x = out_41_cast_fp16)[name = tensor<string, []>("transpose_28")];
            tensor<fp16, [1, 1, 1536]> x_827_cast_fp16 = reshape(shape = var_3660, x = var_3658_cast_fp16)[name = tensor<string, []>("x_827_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403404992)))];
            tensor<fp16, [1, 1, 512]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = x_827_cast_fp16)[name = tensor<string, []>("linear_143_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_829_cast_fp16 = add(x = x_801_cast_fp16, y = linear_143_cast_fp16)[name = tensor<string, []>("x_829_cast_fp16")];
            tensor<string, []> x_829_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_829_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3550_promoted_3 = const()[name = tensor<string, []>("op_3550_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_829_cast_fp16_to_fp32 = cast(dtype = x_829_cast_fp16_to_fp32_dtype_0, x = x_829_cast_fp16)[name = tensor<string, []>("cast_312")];
            tensor<fp32, [1, 1, 512]> var_3671 = pow(x = x_829_cast_fp16_to_fp32, y = var_3550_promoted_3)[name = tensor<string, []>("op_3671")];
            tensor<int32, [1]> var_167_axes_0 = const()[name = tensor<string, []>("var_167_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_167_keep_dims_0 = const()[name = tensor<string, []>("var_167_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_167 = reduce_mean(axes = var_167_axes_0, keep_dims = var_167_keep_dims_0, x = var_3671)[name = tensor<string, []>("var_167")];
            tensor<string, []> var_167_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_167_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3675_to_fp16 = const()[name = tensor<string, []>("op_3675_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_167_to_fp16 = cast(dtype = var_167_to_fp16_dtype_0, x = var_167)[name = tensor<string, []>("cast_311")];
            tensor<fp16, [1, 1, 1]> var_3676_cast_fp16 = add(x = var_167_to_fp16, y = var_3675_to_fp16)[name = tensor<string, []>("op_3676_cast_fp16")];
            tensor<string, []> var_3676_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3676_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3677_epsilon_0 = const()[name = tensor<string, []>("op_3677_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3676_cast_fp16_to_fp32 = cast(dtype = var_3676_cast_fp16_to_fp32_dtype_0, x = var_3676_cast_fp16)[name = tensor<string, []>("cast_310")];
            tensor<fp32, [1, 1, 1]> var_3677 = rsqrt(epsilon = var_3677_epsilon_0, x = var_3676_cast_fp16_to_fp32)[name = tensor<string, []>("op_3677")];
            tensor<string, []> var_3677_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3677_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3677_to_fp16 = cast(dtype = var_3677_to_fp16_dtype_0, x = var_3677)[name = tensor<string, []>("cast_309")];
            tensor<fp16, [1, 1, 512]> x_835_cast_fp16 = mul(x = x_829_cast_fp16, y = var_3677_to_fp16)[name = tensor<string, []>("x_835_cast_fp16")];
            tensor<fp16, [512]> layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404977920)))];
            tensor<fp16, [1, 1, 512]> x_837_cast_fp16 = mul(x = layers_20_post_attention_layernorm_weight_to_fp16, y = x_835_cast_fp16)[name = tensor<string, []>("x_837_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_20_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404979008)))];
            tensor<fp16, [1, 1, 1536]> linear_144_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_20_mlp_gate_proj_weight_to_fp16, x = x_837_cast_fp16)[name = tensor<string, []>("linear_144_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_3688_cast_fp16 = silu(x = linear_144_cast_fp16)[name = tensor<string, []>("op_3688_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_20_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(406551936)))];
            tensor<fp16, [1, 1, 1536]> linear_145_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_20_mlp_up_proj_weight_to_fp16, x = x_837_cast_fp16)[name = tensor<string, []>("linear_145_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_839_cast_fp16 = mul(x = var_3688_cast_fp16, y = linear_145_cast_fp16)[name = tensor<string, []>("x_839_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_20_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_20_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408124864)))];
            tensor<fp16, [1, 1, 512]> linear_146_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_20_mlp_down_proj_weight_to_fp16, x = x_839_cast_fp16)[name = tensor<string, []>("linear_146_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_841_cast_fp16 = add(x = x_829_cast_fp16, y = linear_146_cast_fp16)[name = tensor<string, []>("x_841_cast_fp16")];
            tensor<string, []> x_841_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_841_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_43_begin_0 = const()[name = tensor<string, []>("k_cache_43_begin_0"), val = tensor<int32, [5]>([21, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_43_end_0 = const()[name = tensor<string, []>("k_cache_43_end_0"), val = tensor<int32, [5]>([22, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_43_end_mask_0 = const()[name = tensor<string, []>("k_cache_43_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_43_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_43_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_43_cast_fp16 = slice_by_index(begin = k_cache_43_begin_0, end = k_cache_43_end_0, end_mask = k_cache_43_end_mask_0, squeeze_mask = k_cache_43_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_43_cast_fp16")];
            tensor<int32, [5]> v_cache_43_begin_0 = const()[name = tensor<string, []>("v_cache_43_begin_0"), val = tensor<int32, [5]>([21, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_43_end_0 = const()[name = tensor<string, []>("v_cache_43_end_0"), val = tensor<int32, [5]>([22, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_43_end_mask_0 = const()[name = tensor<string, []>("v_cache_43_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_43_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_43_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_43_cast_fp16 = slice_by_index(begin = v_cache_43_begin_0, end = v_cache_43_end_0, end_mask = v_cache_43_end_mask_0, squeeze_mask = v_cache_43_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_43_cast_fp16")];
            tensor<int32, []> var_3720 = const()[name = tensor<string, []>("op_3720"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_3719_promoted = const()[name = tensor<string, []>("op_3719_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_841_cast_fp16_to_fp32 = cast(dtype = x_841_cast_fp16_to_fp32_dtype_0, x = x_841_cast_fp16)[name = tensor<string, []>("cast_308")];
            tensor<fp32, [1, 1, 512]> var_3729 = pow(x = x_841_cast_fp16_to_fp32, y = var_3719_promoted)[name = tensor<string, []>("op_3729")];
            tensor<int32, [1]> var_169_axes_0 = const()[name = tensor<string, []>("var_169_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_169_keep_dims_0 = const()[name = tensor<string, []>("var_169_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_169 = reduce_mean(axes = var_169_axes_0, keep_dims = var_169_keep_dims_0, x = var_3729)[name = tensor<string, []>("var_169")];
            tensor<string, []> var_169_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_169_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3733_to_fp16 = const()[name = tensor<string, []>("op_3733_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_169_to_fp16 = cast(dtype = var_169_to_fp16_dtype_0, x = var_169)[name = tensor<string, []>("cast_307")];
            tensor<fp16, [1, 1, 1]> var_3734_cast_fp16 = add(x = var_169_to_fp16, y = var_3733_to_fp16)[name = tensor<string, []>("op_3734_cast_fp16")];
            tensor<string, []> var_3734_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3734_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3735_epsilon_0 = const()[name = tensor<string, []>("op_3735_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3734_cast_fp16_to_fp32 = cast(dtype = var_3734_cast_fp16_to_fp32_dtype_0, x = var_3734_cast_fp16)[name = tensor<string, []>("cast_306")];
            tensor<fp32, [1, 1, 1]> var_3735 = rsqrt(epsilon = var_3735_epsilon_0, x = var_3734_cast_fp16_to_fp32)[name = tensor<string, []>("op_3735")];
            tensor<string, []> var_3735_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3735_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3735_to_fp16 = cast(dtype = var_3735_to_fp16_dtype_0, x = var_3735)[name = tensor<string, []>("cast_305")];
            tensor<fp16, [1, 1, 512]> x_847_cast_fp16 = mul(x = x_841_cast_fp16, y = var_3735_to_fp16)[name = tensor<string, []>("x_847_cast_fp16")];
            tensor<fp16, [512]> layers_21_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409697792)))];
            tensor<fp16, [1, 1, 512]> x_849_cast_fp16 = mul(x = layers_21_input_layernorm_weight_to_fp16, y = x_847_cast_fp16)[name = tensor<string, []>("x_849_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(409698880)))];
            tensor<fp16, [1, 1, 1536]> linear_147_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = x_849_cast_fp16)[name = tensor<string, []>("linear_147_cast_fp16")];
            tensor<int32, [4]> var_3749 = const()[name = tensor<string, []>("op_3749"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_851_cast_fp16 = reshape(shape = var_3749, x = linear_147_cast_fp16)[name = tensor<string, []>("x_851_cast_fp16")];
            tensor<string, []> x_851_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_851_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3719_promoted_1 = const()[name = tensor<string, []>("op_3719_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_851_cast_fp16_to_fp32 = cast(dtype = x_851_cast_fp16_to_fp32_dtype_0, x = x_851_cast_fp16)[name = tensor<string, []>("cast_304")];
            tensor<fp32, [1, 1, 12, 128]> var_3753 = pow(x = x_851_cast_fp16_to_fp32, y = var_3719_promoted_1)[name = tensor<string, []>("op_3753")];
            tensor<int32, [1]> var_171_axes_0 = const()[name = tensor<string, []>("var_171_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_171_keep_dims_0 = const()[name = tensor<string, []>("var_171_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_171_0 = reduce_mean(axes = var_171_axes_0, keep_dims = var_171_keep_dims_0, x = var_3753)[name = tensor<string, []>("var_171")];
            tensor<string, []> var_171_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_171_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3757_to_fp16 = const()[name = tensor<string, []>("op_3757_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_171_to_fp16 = cast(dtype = var_171_to_fp16_dtype_0, x = var_171_0)[name = tensor<string, []>("cast_303")];
            tensor<fp16, [1, 1, 12, 1]> var_3758_cast_fp16 = add(x = var_171_to_fp16, y = var_3757_to_fp16)[name = tensor<string, []>("op_3758_cast_fp16")];
            tensor<string, []> var_3758_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3758_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3759_epsilon_0 = const()[name = tensor<string, []>("op_3759_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_3758_cast_fp16_to_fp32 = cast(dtype = var_3758_cast_fp16_to_fp32_dtype_0, x = var_3758_cast_fp16)[name = tensor<string, []>("cast_302")];
            tensor<fp32, [1, 1, 12, 1]> var_3759 = rsqrt(epsilon = var_3759_epsilon_0, x = var_3758_cast_fp16_to_fp32)[name = tensor<string, []>("op_3759")];
            tensor<string, []> var_3759_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3759_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_3759_to_fp16 = cast(dtype = var_3759_to_fp16_dtype_0, x = var_3759)[name = tensor<string, []>("cast_301")];
            tensor<fp16, [1, 1, 12, 128]> x_857_cast_fp16 = mul(x = x_851_cast_fp16, y = var_3759_to_fp16)[name = tensor<string, []>("x_857_cast_fp16")];
            tensor<fp16, [128]> layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411271808)))];
            tensor<fp16, [1, 1, 12, 128]> var_3761_cast_fp16 = mul(x = layers_21_self_attn_q_norm_weight_to_fp16, y = x_857_cast_fp16)[name = tensor<string, []>("op_3761_cast_fp16")];
            tensor<int32, [4]> q_85_perm_0 = const()[name = tensor<string, []>("q_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411272128)))];
            tensor<fp16, [1, 1, 512]> linear_148_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = x_849_cast_fp16)[name = tensor<string, []>("linear_148_cast_fp16")];
            tensor<int32, [4]> var_3765 = const()[name = tensor<string, []>("op_3765"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_859_cast_fp16 = reshape(shape = var_3765, x = linear_148_cast_fp16)[name = tensor<string, []>("x_859_cast_fp16")];
            tensor<string, []> x_859_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_859_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3719_promoted_2 = const()[name = tensor<string, []>("op_3719_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_859_cast_fp16_to_fp32 = cast(dtype = x_859_cast_fp16_to_fp32_dtype_0, x = x_859_cast_fp16)[name = tensor<string, []>("cast_300")];
            tensor<fp32, [1, 1, 4, 128]> var_3769 = pow(x = x_859_cast_fp16_to_fp32, y = var_3719_promoted_2)[name = tensor<string, []>("op_3769")];
            tensor<int32, [1]> var_173_axes_0 = const()[name = tensor<string, []>("var_173_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_173_keep_dims_0 = const()[name = tensor<string, []>("var_173_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_173 = reduce_mean(axes = var_173_axes_0, keep_dims = var_173_keep_dims_0, x = var_3769)[name = tensor<string, []>("var_173")];
            tensor<string, []> var_173_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_173_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3773_to_fp16 = const()[name = tensor<string, []>("op_3773_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_173_to_fp16 = cast(dtype = var_173_to_fp16_dtype_0, x = var_173)[name = tensor<string, []>("cast_299")];
            tensor<fp16, [1, 1, 4, 1]> var_3774_cast_fp16 = add(x = var_173_to_fp16, y = var_3773_to_fp16)[name = tensor<string, []>("op_3774_cast_fp16")];
            tensor<string, []> var_3774_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3774_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3775_epsilon_0 = const()[name = tensor<string, []>("op_3775_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_3774_cast_fp16_to_fp32 = cast(dtype = var_3774_cast_fp16_to_fp32_dtype_0, x = var_3774_cast_fp16)[name = tensor<string, []>("cast_298")];
            tensor<fp32, [1, 1, 4, 1]> var_3775 = rsqrt(epsilon = var_3775_epsilon_0, x = var_3774_cast_fp16_to_fp32)[name = tensor<string, []>("op_3775")];
            tensor<string, []> var_3775_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3775_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_3775_to_fp16 = cast(dtype = var_3775_to_fp16_dtype_0, x = var_3775)[name = tensor<string, []>("cast_297")];
            tensor<fp16, [1, 1, 4, 128]> x_865_cast_fp16 = mul(x = x_859_cast_fp16, y = var_3775_to_fp16)[name = tensor<string, []>("x_865_cast_fp16")];
            tensor<fp16, [128]> layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411796480)))];
            tensor<fp16, [1, 1, 4, 128]> var_3777_cast_fp16 = mul(x = layers_21_self_attn_k_norm_weight_to_fp16, y = x_865_cast_fp16)[name = tensor<string, []>("op_3777_cast_fp16")];
            tensor<int32, [4]> k_85_perm_0 = const()[name = tensor<string, []>("k_85_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(411796800)))];
            tensor<fp16, [1, 1, 512]> linear_149_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = x_849_cast_fp16)[name = tensor<string, []>("linear_149_cast_fp16")];
            tensor<int32, [4]> var_3781 = const()[name = tensor<string, []>("op_3781"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_3782_cast_fp16 = reshape(shape = var_3781, x = linear_149_cast_fp16)[name = tensor<string, []>("op_3782_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = tensor<string, []>("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_85_cast_fp16 = transpose(perm = q_85_perm_0, x = var_3761_cast_fp16)[name = tensor<string, []>("transpose_27")];
            tensor<fp16, [1, 12, 1, 128]> var_3786_cast_fp16 = mul(x = q_85_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3786_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = tensor<string, []>("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = tensor<string, []>("x1_85_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = tensor<string, []>("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_85_cast_fp16)[name = tensor<string, []>("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = tensor<string, []>("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = tensor<string, []>("x2_85_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = tensor<string, []>("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_85_cast_fp16)[name = tensor<string, []>("x2_85_cast_fp16")];
            tensor<fp16, []> const_43_promoted_to_fp16 = const()[name = tensor<string, []>("const_43_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_3789_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_43_promoted_to_fp16)[name = tensor<string, []>("op_3789_cast_fp16")];
            tensor<bool, []> var_3791_interleave_0 = const()[name = tensor<string, []>("op_3791_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_3791_cast_fp16 = concat(axis = var_3720, interleave = var_3791_interleave_0, values = (var_3789_cast_fp16, x1_85_cast_fp16))[name = tensor<string, []>("op_3791_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_3792_cast_fp16 = mul(x = var_3791_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3792_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_87_cast_fp16 = add(x = var_3786_cast_fp16, y = var_3792_cast_fp16)[name = tensor<string, []>("q_87_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_85_cast_fp16 = transpose(perm = k_85_perm_0, x = var_3777_cast_fp16)[name = tensor<string, []>("transpose_26")];
            tensor<fp16, [1, 4, 1, 128]> var_3794_cast_fp16 = mul(x = k_85_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3794_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = tensor<string, []>("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = tensor<string, []>("x1_87_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = tensor<string, []>("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_85_cast_fp16)[name = tensor<string, []>("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = tensor<string, []>("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = tensor<string, []>("x2_87_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = tensor<string, []>("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_85_cast_fp16)[name = tensor<string, []>("x2_87_cast_fp16")];
            tensor<fp16, []> const_44_promoted_to_fp16 = const()[name = tensor<string, []>("const_44_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_3797_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_44_promoted_to_fp16)[name = tensor<string, []>("op_3797_cast_fp16")];
            tensor<bool, []> var_3799_interleave_0 = const()[name = tensor<string, []>("op_3799_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_3799_cast_fp16 = concat(axis = var_3720, interleave = var_3799_interleave_0, values = (var_3797_cast_fp16, x1_87_cast_fp16))[name = tensor<string, []>("op_3799_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_3800_cast_fp16 = mul(x = var_3799_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3800_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_87_cast_fp16 = add(x = var_3794_cast_fp16, y = var_3800_cast_fp16)[name = tensor<string, []>("k_87_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3803_cast_fp16 = mul(x = k_cache_43_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3803_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3804_cast_fp16 = mul(x = k_87_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3804_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_43_cast_fp16 = add(x = var_3803_cast_fp16, y = var_3804_cast_fp16)[name = tensor<string, []>("k_full_43_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3807_cast_fp16 = mul(x = v_cache_43_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3807_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_3782_cast_fp16)[name = tensor<string, []>("transpose_25")];
            tensor<fp16, [1, 4, 2048, 128]> var_3808_cast_fp16 = mul(x = v_43_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3808_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_43_cast_fp16 = add(x = var_3807_cast_fp16, y = var_3808_cast_fp16)[name = tensor<string, []>("v_full_43_cast_fp16")];
            tensor<int32, [1]> var_3810_axes_0 = const()[name = tensor<string, []>("op_3810_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3810_cast_fp16 = expand_dims(axes = var_3810_axes_0, x = k_full_43_cast_fp16)[name = tensor<string, []>("op_3810_cast_fp16")];
            tensor<int32, [5]> var_3812_reps_0 = const()[name = tensor<string, []>("op_3812_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3812_cast_fp16 = tile(reps = var_3812_reps_0, x = var_3810_cast_fp16)[name = tensor<string, []>("op_3812_cast_fp16")];
            tensor<int32, [4]> var_3813 = const()[name = tensor<string, []>("op_3813"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_43_cast_fp16 = reshape(shape = var_3813, x = var_3812_cast_fp16)[name = tensor<string, []>("k_rep_43_cast_fp16")];
            tensor<int32, [1]> var_3815_axes_0 = const()[name = tensor<string, []>("op_3815_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3815_cast_fp16 = expand_dims(axes = var_3815_axes_0, x = v_full_43_cast_fp16)[name = tensor<string, []>("op_3815_cast_fp16")];
            tensor<int32, [5]> var_3817_reps_0 = const()[name = tensor<string, []>("op_3817_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3817_cast_fp16 = tile(reps = var_3817_reps_0, x = var_3815_cast_fp16)[name = tensor<string, []>("op_3817_cast_fp16")];
            tensor<int32, [4]> var_3818 = const()[name = tensor<string, []>("op_3818"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_43_cast_fp16 = reshape(shape = var_3818, x = var_3817_cast_fp16)[name = tensor<string, []>("v_rep_43_cast_fp16")];
            tensor<bool, []> var_3821_transpose_x_1 = const()[name = tensor<string, []>("op_3821_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3821_transpose_y_1 = const()[name = tensor<string, []>("op_3821_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_3821_cast_fp16 = matmul(transpose_x = var_3821_transpose_x_1, transpose_y = var_3821_transpose_y_1, x = q_87_cast_fp16, y = k_rep_43_cast_fp16)[name = tensor<string, []>("op_3821_cast_fp16")];
            tensor<fp16, []> var_3822_to_fp16 = const()[name = tensor<string, []>("op_3822_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_85_cast_fp16 = mul(x = var_3821_cast_fp16, y = var_3822_to_fp16)[name = tensor<string, []>("attn_85_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_87_cast_fp16 = add(x = attn_85_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_87_cast_fp16")];
            tensor<string, []> input_87_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_87_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_87_cast_fp16_to_fp32 = cast(dtype = input_87_cast_fp16_to_fp32_dtype_0, x = input_87_cast_fp16)[name = tensor<string, []>("cast_296")];
            tensor<fp32, [1, 12, 1, 2048]> attn_87 = softmax(axis = var_3720, x = input_87_cast_fp16_to_fp32)[name = tensor<string, []>("attn_87")];
            tensor<bool, []> out_43_transpose_x_0 = const()[name = tensor<string, []>("out_43_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_43_transpose_y_0 = const()[name = tensor<string, []>("out_43_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_87_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_87_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_87_to_fp16 = cast(dtype = attn_87_to_fp16_dtype_0, x = attn_87)[name = tensor<string, []>("cast_295")];
            tensor<fp16, [1, 12, 1, 128]> out_43_cast_fp16 = matmul(transpose_x = out_43_transpose_x_0, transpose_y = out_43_transpose_y_0, x = attn_87_to_fp16, y = v_rep_43_cast_fp16)[name = tensor<string, []>("out_43_cast_fp16")];
            tensor<int32, [4]> var_3827_perm_0 = const()[name = tensor<string, []>("op_3827_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3829 = const()[name = tensor<string, []>("op_3829"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_3827_cast_fp16 = transpose(perm = var_3827_perm_0, x = out_43_cast_fp16)[name = tensor<string, []>("transpose_24")];
            tensor<fp16, [1, 1, 1536]> x_867_cast_fp16 = reshape(shape = var_3829, x = var_3827_cast_fp16)[name = tensor<string, []>("x_867_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(412321152)))];
            tensor<fp16, [1, 1, 512]> linear_150_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = x_867_cast_fp16)[name = tensor<string, []>("linear_150_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_869_cast_fp16 = add(x = x_841_cast_fp16, y = linear_150_cast_fp16)[name = tensor<string, []>("x_869_cast_fp16")];
            tensor<string, []> x_869_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_869_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3719_promoted_3 = const()[name = tensor<string, []>("op_3719_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_869_cast_fp16_to_fp32 = cast(dtype = x_869_cast_fp16_to_fp32_dtype_0, x = x_869_cast_fp16)[name = tensor<string, []>("cast_294")];
            tensor<fp32, [1, 1, 512]> var_3840 = pow(x = x_869_cast_fp16_to_fp32, y = var_3719_promoted_3)[name = tensor<string, []>("op_3840")];
            tensor<int32, [1]> var_175_axes_0 = const()[name = tensor<string, []>("var_175_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_175_keep_dims_0 = const()[name = tensor<string, []>("var_175_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_175 = reduce_mean(axes = var_175_axes_0, keep_dims = var_175_keep_dims_0, x = var_3840)[name = tensor<string, []>("var_175")];
            tensor<string, []> var_175_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_175_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3844_to_fp16 = const()[name = tensor<string, []>("op_3844_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_175_to_fp16 = cast(dtype = var_175_to_fp16_dtype_0, x = var_175)[name = tensor<string, []>("cast_293")];
            tensor<fp16, [1, 1, 1]> var_3845_cast_fp16 = add(x = var_175_to_fp16, y = var_3844_to_fp16)[name = tensor<string, []>("op_3845_cast_fp16")];
            tensor<string, []> var_3845_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3845_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3846_epsilon_0 = const()[name = tensor<string, []>("op_3846_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3845_cast_fp16_to_fp32 = cast(dtype = var_3845_cast_fp16_to_fp32_dtype_0, x = var_3845_cast_fp16)[name = tensor<string, []>("cast_292")];
            tensor<fp32, [1, 1, 1]> var_3846 = rsqrt(epsilon = var_3846_epsilon_0, x = var_3845_cast_fp16_to_fp32)[name = tensor<string, []>("op_3846")];
            tensor<string, []> var_3846_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3846_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3846_to_fp16 = cast(dtype = var_3846_to_fp16_dtype_0, x = var_3846)[name = tensor<string, []>("cast_291")];
            tensor<fp16, [1, 1, 512]> x_875_cast_fp16 = mul(x = x_869_cast_fp16, y = var_3846_to_fp16)[name = tensor<string, []>("x_875_cast_fp16")];
            tensor<fp16, [512]> layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(413894080)))];
            tensor<fp16, [1, 1, 512]> x_877_cast_fp16 = mul(x = layers_21_post_attention_layernorm_weight_to_fp16, y = x_875_cast_fp16)[name = tensor<string, []>("x_877_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_21_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(413895168)))];
            tensor<fp16, [1, 1, 1536]> linear_151_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_21_mlp_gate_proj_weight_to_fp16, x = x_877_cast_fp16)[name = tensor<string, []>("linear_151_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_3857_cast_fp16 = silu(x = linear_151_cast_fp16)[name = tensor<string, []>("op_3857_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_21_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(415468096)))];
            tensor<fp16, [1, 1, 1536]> linear_152_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_21_mlp_up_proj_weight_to_fp16, x = x_877_cast_fp16)[name = tensor<string, []>("linear_152_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_879_cast_fp16 = mul(x = var_3857_cast_fp16, y = linear_152_cast_fp16)[name = tensor<string, []>("x_879_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_21_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_21_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417041024)))];
            tensor<fp16, [1, 1, 512]> linear_153_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_21_mlp_down_proj_weight_to_fp16, x = x_879_cast_fp16)[name = tensor<string, []>("linear_153_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_881_cast_fp16 = add(x = x_869_cast_fp16, y = linear_153_cast_fp16)[name = tensor<string, []>("x_881_cast_fp16")];
            tensor<string, []> x_881_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_881_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_45_begin_0 = const()[name = tensor<string, []>("k_cache_45_begin_0"), val = tensor<int32, [5]>([22, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_45_end_0 = const()[name = tensor<string, []>("k_cache_45_end_0"), val = tensor<int32, [5]>([23, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_45_end_mask_0 = const()[name = tensor<string, []>("k_cache_45_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_45_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_45_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_45_cast_fp16 = slice_by_index(begin = k_cache_45_begin_0, end = k_cache_45_end_0, end_mask = k_cache_45_end_mask_0, squeeze_mask = k_cache_45_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_45_cast_fp16")];
            tensor<int32, [5]> v_cache_45_begin_0 = const()[name = tensor<string, []>("v_cache_45_begin_0"), val = tensor<int32, [5]>([22, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_45_end_0 = const()[name = tensor<string, []>("v_cache_45_end_0"), val = tensor<int32, [5]>([23, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_45_end_mask_0 = const()[name = tensor<string, []>("v_cache_45_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_45_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_45_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_45_cast_fp16 = slice_by_index(begin = v_cache_45_begin_0, end = v_cache_45_end_0, end_mask = v_cache_45_end_mask_0, squeeze_mask = v_cache_45_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_45_cast_fp16")];
            tensor<int32, []> var_3889 = const()[name = tensor<string, []>("op_3889"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_3888_promoted = const()[name = tensor<string, []>("op_3888_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_881_cast_fp16_to_fp32 = cast(dtype = x_881_cast_fp16_to_fp32_dtype_0, x = x_881_cast_fp16)[name = tensor<string, []>("cast_290")];
            tensor<fp32, [1, 1, 512]> var_3898 = pow(x = x_881_cast_fp16_to_fp32, y = var_3888_promoted)[name = tensor<string, []>("op_3898")];
            tensor<int32, [1]> var_177_axes_0 = const()[name = tensor<string, []>("var_177_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_177_keep_dims_0 = const()[name = tensor<string, []>("var_177_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_177 = reduce_mean(axes = var_177_axes_0, keep_dims = var_177_keep_dims_0, x = var_3898)[name = tensor<string, []>("var_177")];
            tensor<string, []> var_177_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_177_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3902_to_fp16 = const()[name = tensor<string, []>("op_3902_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_177_to_fp16 = cast(dtype = var_177_to_fp16_dtype_0, x = var_177)[name = tensor<string, []>("cast_289")];
            tensor<fp16, [1, 1, 1]> var_3903_cast_fp16 = add(x = var_177_to_fp16, y = var_3902_to_fp16)[name = tensor<string, []>("op_3903_cast_fp16")];
            tensor<string, []> var_3903_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3903_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3904_epsilon_0 = const()[name = tensor<string, []>("op_3904_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_3903_cast_fp16_to_fp32 = cast(dtype = var_3903_cast_fp16_to_fp32_dtype_0, x = var_3903_cast_fp16)[name = tensor<string, []>("cast_288")];
            tensor<fp32, [1, 1, 1]> var_3904 = rsqrt(epsilon = var_3904_epsilon_0, x = var_3903_cast_fp16_to_fp32)[name = tensor<string, []>("op_3904")];
            tensor<string, []> var_3904_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3904_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_3904_to_fp16 = cast(dtype = var_3904_to_fp16_dtype_0, x = var_3904)[name = tensor<string, []>("cast_287")];
            tensor<fp16, [1, 1, 512]> x_887_cast_fp16 = mul(x = x_881_cast_fp16, y = var_3904_to_fp16)[name = tensor<string, []>("x_887_cast_fp16")];
            tensor<fp16, [512]> layers_22_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418613952)))];
            tensor<fp16, [1, 1, 512]> x_889_cast_fp16 = mul(x = layers_22_input_layernorm_weight_to_fp16, y = x_887_cast_fp16)[name = tensor<string, []>("x_889_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(418615040)))];
            tensor<fp16, [1, 1, 1536]> linear_154_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = x_889_cast_fp16)[name = tensor<string, []>("linear_154_cast_fp16")];
            tensor<int32, [4]> var_3918 = const()[name = tensor<string, []>("op_3918"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_891_cast_fp16 = reshape(shape = var_3918, x = linear_154_cast_fp16)[name = tensor<string, []>("x_891_cast_fp16")];
            tensor<string, []> x_891_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_891_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3888_promoted_1 = const()[name = tensor<string, []>("op_3888_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_891_cast_fp16_to_fp32 = cast(dtype = x_891_cast_fp16_to_fp32_dtype_0, x = x_891_cast_fp16)[name = tensor<string, []>("cast_286")];
            tensor<fp32, [1, 1, 12, 128]> var_3922 = pow(x = x_891_cast_fp16_to_fp32, y = var_3888_promoted_1)[name = tensor<string, []>("op_3922")];
            tensor<int32, [1]> var_179_axes_0 = const()[name = tensor<string, []>("var_179_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_179_keep_dims_0 = const()[name = tensor<string, []>("var_179_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_179 = reduce_mean(axes = var_179_axes_0, keep_dims = var_179_keep_dims_0, x = var_3922)[name = tensor<string, []>("var_179")];
            tensor<string, []> var_179_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_179_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3926_to_fp16 = const()[name = tensor<string, []>("op_3926_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_179_to_fp16 = cast(dtype = var_179_to_fp16_dtype_0, x = var_179)[name = tensor<string, []>("cast_285")];
            tensor<fp16, [1, 1, 12, 1]> var_3927_cast_fp16 = add(x = var_179_to_fp16, y = var_3926_to_fp16)[name = tensor<string, []>("op_3927_cast_fp16")];
            tensor<string, []> var_3927_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3927_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3928_epsilon_0 = const()[name = tensor<string, []>("op_3928_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_3927_cast_fp16_to_fp32 = cast(dtype = var_3927_cast_fp16_to_fp32_dtype_0, x = var_3927_cast_fp16)[name = tensor<string, []>("cast_284")];
            tensor<fp32, [1, 1, 12, 1]> var_3928 = rsqrt(epsilon = var_3928_epsilon_0, x = var_3927_cast_fp16_to_fp32)[name = tensor<string, []>("op_3928")];
            tensor<string, []> var_3928_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3928_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_3928_to_fp16 = cast(dtype = var_3928_to_fp16_dtype_0, x = var_3928)[name = tensor<string, []>("cast_283")];
            tensor<fp16, [1, 1, 12, 128]> x_897_cast_fp16 = mul(x = x_891_cast_fp16, y = var_3928_to_fp16)[name = tensor<string, []>("x_897_cast_fp16")];
            tensor<fp16, [128]> layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420187968)))];
            tensor<fp16, [1, 1, 12, 128]> var_3930_cast_fp16 = mul(x = layers_22_self_attn_q_norm_weight_to_fp16, y = x_897_cast_fp16)[name = tensor<string, []>("op_3930_cast_fp16")];
            tensor<int32, [4]> q_89_perm_0 = const()[name = tensor<string, []>("q_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420188288)))];
            tensor<fp16, [1, 1, 512]> linear_155_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = x_889_cast_fp16)[name = tensor<string, []>("linear_155_cast_fp16")];
            tensor<int32, [4]> var_3934 = const()[name = tensor<string, []>("op_3934"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_899_cast_fp16 = reshape(shape = var_3934, x = linear_155_cast_fp16)[name = tensor<string, []>("x_899_cast_fp16")];
            tensor<string, []> x_899_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_899_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3888_promoted_2 = const()[name = tensor<string, []>("op_3888_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_899_cast_fp16_to_fp32 = cast(dtype = x_899_cast_fp16_to_fp32_dtype_0, x = x_899_cast_fp16)[name = tensor<string, []>("cast_282")];
            tensor<fp32, [1, 1, 4, 128]> var_3938 = pow(x = x_899_cast_fp16_to_fp32, y = var_3888_promoted_2)[name = tensor<string, []>("op_3938")];
            tensor<int32, [1]> var_181_axes_0 = const()[name = tensor<string, []>("var_181_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_181_keep_dims_0 = const()[name = tensor<string, []>("var_181_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_181 = reduce_mean(axes = var_181_axes_0, keep_dims = var_181_keep_dims_0, x = var_3938)[name = tensor<string, []>("var_181")];
            tensor<string, []> var_181_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_181_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_3942_to_fp16 = const()[name = tensor<string, []>("op_3942_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_181_to_fp16 = cast(dtype = var_181_to_fp16_dtype_0, x = var_181)[name = tensor<string, []>("cast_281")];
            tensor<fp16, [1, 1, 4, 1]> var_3943_cast_fp16 = add(x = var_181_to_fp16, y = var_3942_to_fp16)[name = tensor<string, []>("op_3943_cast_fp16")];
            tensor<string, []> var_3943_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_3943_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3944_epsilon_0 = const()[name = tensor<string, []>("op_3944_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_3943_cast_fp16_to_fp32 = cast(dtype = var_3943_cast_fp16_to_fp32_dtype_0, x = var_3943_cast_fp16)[name = tensor<string, []>("cast_280")];
            tensor<fp32, [1, 1, 4, 1]> var_3944 = rsqrt(epsilon = var_3944_epsilon_0, x = var_3943_cast_fp16_to_fp32)[name = tensor<string, []>("op_3944")];
            tensor<string, []> var_3944_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_3944_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_3944_to_fp16 = cast(dtype = var_3944_to_fp16_dtype_0, x = var_3944)[name = tensor<string, []>("cast_279")];
            tensor<fp16, [1, 1, 4, 128]> x_905_cast_fp16 = mul(x = x_899_cast_fp16, y = var_3944_to_fp16)[name = tensor<string, []>("x_905_cast_fp16")];
            tensor<fp16, [128]> layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420712640)))];
            tensor<fp16, [1, 1, 4, 128]> var_3946_cast_fp16 = mul(x = layers_22_self_attn_k_norm_weight_to_fp16, y = x_905_cast_fp16)[name = tensor<string, []>("op_3946_cast_fp16")];
            tensor<int32, [4]> k_89_perm_0 = const()[name = tensor<string, []>("k_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(420712960)))];
            tensor<fp16, [1, 1, 512]> linear_156_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = x_889_cast_fp16)[name = tensor<string, []>("linear_156_cast_fp16")];
            tensor<int32, [4]> var_3950 = const()[name = tensor<string, []>("op_3950"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_3951_cast_fp16 = reshape(shape = var_3950, x = linear_156_cast_fp16)[name = tensor<string, []>("op_3951_cast_fp16")];
            tensor<int32, [4]> v_45_perm_0 = const()[name = tensor<string, []>("v_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_89_cast_fp16 = transpose(perm = q_89_perm_0, x = var_3930_cast_fp16)[name = tensor<string, []>("transpose_23")];
            tensor<fp16, [1, 12, 1, 128]> var_3955_cast_fp16 = mul(x = q_89_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3955_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = tensor<string, []>("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = tensor<string, []>("x1_89_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = tensor<string, []>("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_89_cast_fp16)[name = tensor<string, []>("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = tensor<string, []>("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = tensor<string, []>("x2_89_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = tensor<string, []>("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_89_cast_fp16)[name = tensor<string, []>("x2_89_cast_fp16")];
            tensor<fp16, []> const_45_promoted_to_fp16 = const()[name = tensor<string, []>("const_45_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_3958_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_45_promoted_to_fp16)[name = tensor<string, []>("op_3958_cast_fp16")];
            tensor<bool, []> var_3960_interleave_0 = const()[name = tensor<string, []>("op_3960_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_3960_cast_fp16 = concat(axis = var_3889, interleave = var_3960_interleave_0, values = (var_3958_cast_fp16, x1_89_cast_fp16))[name = tensor<string, []>("op_3960_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_3961_cast_fp16 = mul(x = var_3960_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3961_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_91_cast_fp16 = add(x = var_3955_cast_fp16, y = var_3961_cast_fp16)[name = tensor<string, []>("q_91_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_89_cast_fp16 = transpose(perm = k_89_perm_0, x = var_3946_cast_fp16)[name = tensor<string, []>("transpose_22")];
            tensor<fp16, [1, 4, 1, 128]> var_3963_cast_fp16 = mul(x = k_89_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_3963_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = tensor<string, []>("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = tensor<string, []>("x1_91_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = tensor<string, []>("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_89_cast_fp16)[name = tensor<string, []>("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = tensor<string, []>("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = tensor<string, []>("x2_91_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = tensor<string, []>("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_89_cast_fp16)[name = tensor<string, []>("x2_91_cast_fp16")];
            tensor<fp16, []> const_46_promoted_to_fp16 = const()[name = tensor<string, []>("const_46_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_3966_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_46_promoted_to_fp16)[name = tensor<string, []>("op_3966_cast_fp16")];
            tensor<bool, []> var_3968_interleave_0 = const()[name = tensor<string, []>("op_3968_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_3968_cast_fp16 = concat(axis = var_3889, interleave = var_3968_interleave_0, values = (var_3966_cast_fp16, x1_91_cast_fp16))[name = tensor<string, []>("op_3968_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_3969_cast_fp16 = mul(x = var_3968_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_3969_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_91_cast_fp16 = add(x = var_3963_cast_fp16, y = var_3969_cast_fp16)[name = tensor<string, []>("k_91_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3972_cast_fp16 = mul(x = k_cache_45_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3972_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3973_cast_fp16 = mul(x = k_91_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3973_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_45_cast_fp16 = add(x = var_3972_cast_fp16, y = var_3973_cast_fp16)[name = tensor<string, []>("k_full_45_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_3976_cast_fp16 = mul(x = v_cache_45_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_3976_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_3951_cast_fp16)[name = tensor<string, []>("transpose_21")];
            tensor<fp16, [1, 4, 2048, 128]> var_3977_cast_fp16 = mul(x = v_45_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_3977_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_45_cast_fp16 = add(x = var_3976_cast_fp16, y = var_3977_cast_fp16)[name = tensor<string, []>("v_full_45_cast_fp16")];
            tensor<int32, [1]> var_3979_axes_0 = const()[name = tensor<string, []>("op_3979_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3979_cast_fp16 = expand_dims(axes = var_3979_axes_0, x = k_full_45_cast_fp16)[name = tensor<string, []>("op_3979_cast_fp16")];
            tensor<int32, [5]> var_3981_reps_0 = const()[name = tensor<string, []>("op_3981_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3981_cast_fp16 = tile(reps = var_3981_reps_0, x = var_3979_cast_fp16)[name = tensor<string, []>("op_3981_cast_fp16")];
            tensor<int32, [4]> var_3982 = const()[name = tensor<string, []>("op_3982"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_45_cast_fp16 = reshape(shape = var_3982, x = var_3981_cast_fp16)[name = tensor<string, []>("k_rep_45_cast_fp16")];
            tensor<int32, [1]> var_3984_axes_0 = const()[name = tensor<string, []>("op_3984_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_3984_cast_fp16 = expand_dims(axes = var_3984_axes_0, x = v_full_45_cast_fp16)[name = tensor<string, []>("op_3984_cast_fp16")];
            tensor<int32, [5]> var_3986_reps_0 = const()[name = tensor<string, []>("op_3986_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_3986_cast_fp16 = tile(reps = var_3986_reps_0, x = var_3984_cast_fp16)[name = tensor<string, []>("op_3986_cast_fp16")];
            tensor<int32, [4]> var_3987 = const()[name = tensor<string, []>("op_3987"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_45_cast_fp16 = reshape(shape = var_3987, x = var_3986_cast_fp16)[name = tensor<string, []>("v_rep_45_cast_fp16")];
            tensor<bool, []> var_3990_transpose_x_1 = const()[name = tensor<string, []>("op_3990_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_3990_transpose_y_1 = const()[name = tensor<string, []>("op_3990_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_3990_cast_fp16 = matmul(transpose_x = var_3990_transpose_x_1, transpose_y = var_3990_transpose_y_1, x = q_91_cast_fp16, y = k_rep_45_cast_fp16)[name = tensor<string, []>("op_3990_cast_fp16")];
            tensor<fp16, []> var_3991_to_fp16 = const()[name = tensor<string, []>("op_3991_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_89_cast_fp16 = mul(x = var_3990_cast_fp16, y = var_3991_to_fp16)[name = tensor<string, []>("attn_89_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_91_cast_fp16 = add(x = attn_89_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_91_cast_fp16")];
            tensor<string, []> input_91_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_91_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_91_cast_fp16_to_fp32 = cast(dtype = input_91_cast_fp16_to_fp32_dtype_0, x = input_91_cast_fp16)[name = tensor<string, []>("cast_278")];
            tensor<fp32, [1, 12, 1, 2048]> attn_91 = softmax(axis = var_3889, x = input_91_cast_fp16_to_fp32)[name = tensor<string, []>("attn_91")];
            tensor<bool, []> out_45_transpose_x_0 = const()[name = tensor<string, []>("out_45_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_45_transpose_y_0 = const()[name = tensor<string, []>("out_45_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_91_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_91_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_91_to_fp16 = cast(dtype = attn_91_to_fp16_dtype_0, x = attn_91)[name = tensor<string, []>("cast_277")];
            tensor<fp16, [1, 12, 1, 128]> out_45_cast_fp16 = matmul(transpose_x = out_45_transpose_x_0, transpose_y = out_45_transpose_y_0, x = attn_91_to_fp16, y = v_rep_45_cast_fp16)[name = tensor<string, []>("out_45_cast_fp16")];
            tensor<int32, [4]> var_3996_perm_0 = const()[name = tensor<string, []>("op_3996_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3998 = const()[name = tensor<string, []>("op_3998"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_3996_cast_fp16 = transpose(perm = var_3996_perm_0, x = out_45_cast_fp16)[name = tensor<string, []>("transpose_20")];
            tensor<fp16, [1, 1, 1536]> x_907_cast_fp16 = reshape(shape = var_3998, x = var_3996_cast_fp16)[name = tensor<string, []>("x_907_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421237312)))];
            tensor<fp16, [1, 1, 512]> linear_157_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = x_907_cast_fp16)[name = tensor<string, []>("linear_157_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_909_cast_fp16 = add(x = x_881_cast_fp16, y = linear_157_cast_fp16)[name = tensor<string, []>("x_909_cast_fp16")];
            tensor<string, []> x_909_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_909_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_3888_promoted_3 = const()[name = tensor<string, []>("op_3888_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_909_cast_fp16_to_fp32 = cast(dtype = x_909_cast_fp16_to_fp32_dtype_0, x = x_909_cast_fp16)[name = tensor<string, []>("cast_276")];
            tensor<fp32, [1, 1, 512]> var_4009 = pow(x = x_909_cast_fp16_to_fp32, y = var_3888_promoted_3)[name = tensor<string, []>("op_4009")];
            tensor<int32, [1]> var_183_axes_0 = const()[name = tensor<string, []>("var_183_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_183_keep_dims_0 = const()[name = tensor<string, []>("var_183_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_183 = reduce_mean(axes = var_183_axes_0, keep_dims = var_183_keep_dims_0, x = var_4009)[name = tensor<string, []>("var_183")];
            tensor<string, []> var_183_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_183_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4013_to_fp16 = const()[name = tensor<string, []>("op_4013_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_183_to_fp16 = cast(dtype = var_183_to_fp16_dtype_0, x = var_183)[name = tensor<string, []>("cast_275")];
            tensor<fp16, [1, 1, 1]> var_4014_cast_fp16 = add(x = var_183_to_fp16, y = var_4013_to_fp16)[name = tensor<string, []>("op_4014_cast_fp16")];
            tensor<string, []> var_4014_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4014_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4015_epsilon_0 = const()[name = tensor<string, []>("op_4015_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4014_cast_fp16_to_fp32 = cast(dtype = var_4014_cast_fp16_to_fp32_dtype_0, x = var_4014_cast_fp16)[name = tensor<string, []>("cast_274")];
            tensor<fp32, [1, 1, 1]> var_4015 = rsqrt(epsilon = var_4015_epsilon_0, x = var_4014_cast_fp16_to_fp32)[name = tensor<string, []>("op_4015")];
            tensor<string, []> var_4015_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4015_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4015_to_fp16 = cast(dtype = var_4015_to_fp16_dtype_0, x = var_4015)[name = tensor<string, []>("cast_273")];
            tensor<fp16, [1, 1, 512]> x_915_cast_fp16 = mul(x = x_909_cast_fp16, y = var_4015_to_fp16)[name = tensor<string, []>("x_915_cast_fp16")];
            tensor<fp16, [512]> layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422810240)))];
            tensor<fp16, [1, 1, 512]> x_917_cast_fp16 = mul(x = layers_22_post_attention_layernorm_weight_to_fp16, y = x_915_cast_fp16)[name = tensor<string, []>("x_917_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_22_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(422811328)))];
            tensor<fp16, [1, 1, 1536]> linear_158_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_22_mlp_gate_proj_weight_to_fp16, x = x_917_cast_fp16)[name = tensor<string, []>("linear_158_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_4026_cast_fp16 = silu(x = linear_158_cast_fp16)[name = tensor<string, []>("op_4026_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_22_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(424384256)))];
            tensor<fp16, [1, 1, 1536]> linear_159_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_22_mlp_up_proj_weight_to_fp16, x = x_917_cast_fp16)[name = tensor<string, []>("linear_159_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_919_cast_fp16 = mul(x = var_4026_cast_fp16, y = linear_159_cast_fp16)[name = tensor<string, []>("x_919_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_22_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_22_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425957184)))];
            tensor<fp16, [1, 1, 512]> linear_160_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_22_mlp_down_proj_weight_to_fp16, x = x_919_cast_fp16)[name = tensor<string, []>("linear_160_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_921_cast_fp16 = add(x = x_909_cast_fp16, y = linear_160_cast_fp16)[name = tensor<string, []>("x_921_cast_fp16")];
            tensor<string, []> x_921_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_921_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_47_begin_0 = const()[name = tensor<string, []>("k_cache_47_begin_0"), val = tensor<int32, [5]>([23, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_47_end_0 = const()[name = tensor<string, []>("k_cache_47_end_0"), val = tensor<int32, [5]>([24, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_47_end_mask_0 = const()[name = tensor<string, []>("k_cache_47_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_47_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_47_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_47_cast_fp16 = slice_by_index(begin = k_cache_47_begin_0, end = k_cache_47_end_0, end_mask = k_cache_47_end_mask_0, squeeze_mask = k_cache_47_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_47_cast_fp16")];
            tensor<int32, [5]> v_cache_47_begin_0 = const()[name = tensor<string, []>("v_cache_47_begin_0"), val = tensor<int32, [5]>([23, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_47_end_0 = const()[name = tensor<string, []>("v_cache_47_end_0"), val = tensor<int32, [5]>([24, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_47_end_mask_0 = const()[name = tensor<string, []>("v_cache_47_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_47_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_47_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_47_cast_fp16 = slice_by_index(begin = v_cache_47_begin_0, end = v_cache_47_end_0, end_mask = v_cache_47_end_mask_0, squeeze_mask = v_cache_47_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_47_cast_fp16")];
            tensor<int32, []> var_4058 = const()[name = tensor<string, []>("op_4058"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_4057_promoted = const()[name = tensor<string, []>("op_4057_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_921_cast_fp16_to_fp32 = cast(dtype = x_921_cast_fp16_to_fp32_dtype_0, x = x_921_cast_fp16)[name = tensor<string, []>("cast_272")];
            tensor<fp32, [1, 1, 512]> var_4067 = pow(x = x_921_cast_fp16_to_fp32, y = var_4057_promoted)[name = tensor<string, []>("op_4067")];
            tensor<int32, [1]> var_185_axes_0 = const()[name = tensor<string, []>("var_185_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_185_keep_dims_0 = const()[name = tensor<string, []>("var_185_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_185 = reduce_mean(axes = var_185_axes_0, keep_dims = var_185_keep_dims_0, x = var_4067)[name = tensor<string, []>("var_185")];
            tensor<string, []> var_185_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_185_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4071_to_fp16 = const()[name = tensor<string, []>("op_4071_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_185_to_fp16 = cast(dtype = var_185_to_fp16_dtype_0, x = var_185)[name = tensor<string, []>("cast_271")];
            tensor<fp16, [1, 1, 1]> var_4072_cast_fp16 = add(x = var_185_to_fp16, y = var_4071_to_fp16)[name = tensor<string, []>("op_4072_cast_fp16")];
            tensor<string, []> var_4072_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4072_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4073_epsilon_0 = const()[name = tensor<string, []>("op_4073_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4072_cast_fp16_to_fp32 = cast(dtype = var_4072_cast_fp16_to_fp32_dtype_0, x = var_4072_cast_fp16)[name = tensor<string, []>("cast_270")];
            tensor<fp32, [1, 1, 1]> var_4073 = rsqrt(epsilon = var_4073_epsilon_0, x = var_4072_cast_fp16_to_fp32)[name = tensor<string, []>("op_4073")];
            tensor<string, []> var_4073_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4073_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4073_to_fp16 = cast(dtype = var_4073_to_fp16_dtype_0, x = var_4073)[name = tensor<string, []>("cast_269")];
            tensor<fp16, [1, 1, 512]> x_927_cast_fp16 = mul(x = x_921_cast_fp16, y = var_4073_to_fp16)[name = tensor<string, []>("x_927_cast_fp16")];
            tensor<fp16, [512]> layers_23_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(427530112)))];
            tensor<fp16, [1, 1, 512]> x_929_cast_fp16 = mul(x = layers_23_input_layernorm_weight_to_fp16, y = x_927_cast_fp16)[name = tensor<string, []>("x_929_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(427531200)))];
            tensor<fp16, [1, 1, 1536]> linear_161_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = x_929_cast_fp16)[name = tensor<string, []>("linear_161_cast_fp16")];
            tensor<int32, [4]> var_4087 = const()[name = tensor<string, []>("op_4087"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_931_cast_fp16 = reshape(shape = var_4087, x = linear_161_cast_fp16)[name = tensor<string, []>("x_931_cast_fp16")];
            tensor<string, []> x_931_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_931_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4057_promoted_1 = const()[name = tensor<string, []>("op_4057_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_931_cast_fp16_to_fp32 = cast(dtype = x_931_cast_fp16_to_fp32_dtype_0, x = x_931_cast_fp16)[name = tensor<string, []>("cast_268")];
            tensor<fp32, [1, 1, 12, 128]> var_4091 = pow(x = x_931_cast_fp16_to_fp32, y = var_4057_promoted_1)[name = tensor<string, []>("op_4091")];
            tensor<int32, [1]> var_187_axes_0 = const()[name = tensor<string, []>("var_187_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_187_keep_dims_0 = const()[name = tensor<string, []>("var_187_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_187 = reduce_mean(axes = var_187_axes_0, keep_dims = var_187_keep_dims_0, x = var_4091)[name = tensor<string, []>("var_187")];
            tensor<string, []> var_187_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_187_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4095_to_fp16 = const()[name = tensor<string, []>("op_4095_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_187_to_fp16 = cast(dtype = var_187_to_fp16_dtype_0, x = var_187)[name = tensor<string, []>("cast_267")];
            tensor<fp16, [1, 1, 12, 1]> var_4096_cast_fp16 = add(x = var_187_to_fp16, y = var_4095_to_fp16)[name = tensor<string, []>("op_4096_cast_fp16")];
            tensor<string, []> var_4096_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4096_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4097_epsilon_0 = const()[name = tensor<string, []>("op_4097_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_4096_cast_fp16_to_fp32 = cast(dtype = var_4096_cast_fp16_to_fp32_dtype_0, x = var_4096_cast_fp16)[name = tensor<string, []>("cast_266")];
            tensor<fp32, [1, 1, 12, 1]> var_4097 = rsqrt(epsilon = var_4097_epsilon_0, x = var_4096_cast_fp16_to_fp32)[name = tensor<string, []>("op_4097")];
            tensor<string, []> var_4097_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4097_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_4097_to_fp16 = cast(dtype = var_4097_to_fp16_dtype_0, x = var_4097)[name = tensor<string, []>("cast_265")];
            tensor<fp16, [1, 1, 12, 128]> x_937_cast_fp16 = mul(x = x_931_cast_fp16, y = var_4097_to_fp16)[name = tensor<string, []>("x_937_cast_fp16")];
            tensor<fp16, [128]> layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429104128)))];
            tensor<fp16, [1, 1, 12, 128]> var_4099_cast_fp16 = mul(x = layers_23_self_attn_q_norm_weight_to_fp16, y = x_937_cast_fp16)[name = tensor<string, []>("op_4099_cast_fp16")];
            tensor<int32, [4]> q_93_perm_0 = const()[name = tensor<string, []>("q_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429104448)))];
            tensor<fp16, [1, 1, 512]> linear_162_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = x_929_cast_fp16)[name = tensor<string, []>("linear_162_cast_fp16")];
            tensor<int32, [4]> var_4103 = const()[name = tensor<string, []>("op_4103"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_939_cast_fp16 = reshape(shape = var_4103, x = linear_162_cast_fp16)[name = tensor<string, []>("x_939_cast_fp16")];
            tensor<string, []> x_939_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_939_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4057_promoted_2 = const()[name = tensor<string, []>("op_4057_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_939_cast_fp16_to_fp32 = cast(dtype = x_939_cast_fp16_to_fp32_dtype_0, x = x_939_cast_fp16)[name = tensor<string, []>("cast_264")];
            tensor<fp32, [1, 1, 4, 128]> var_4107 = pow(x = x_939_cast_fp16_to_fp32, y = var_4057_promoted_2)[name = tensor<string, []>("op_4107")];
            tensor<int32, [1]> var_189_axes_0 = const()[name = tensor<string, []>("var_189_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_189_keep_dims_0 = const()[name = tensor<string, []>("var_189_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_189 = reduce_mean(axes = var_189_axes_0, keep_dims = var_189_keep_dims_0, x = var_4107)[name = tensor<string, []>("var_189")];
            tensor<string, []> var_189_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_189_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4111_to_fp16 = const()[name = tensor<string, []>("op_4111_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_189_to_fp16 = cast(dtype = var_189_to_fp16_dtype_0, x = var_189)[name = tensor<string, []>("cast_263")];
            tensor<fp16, [1, 1, 4, 1]> var_4112_cast_fp16 = add(x = var_189_to_fp16, y = var_4111_to_fp16)[name = tensor<string, []>("op_4112_cast_fp16")];
            tensor<string, []> var_4112_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4112_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4113_epsilon_0 = const()[name = tensor<string, []>("op_4113_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_4112_cast_fp16_to_fp32 = cast(dtype = var_4112_cast_fp16_to_fp32_dtype_0, x = var_4112_cast_fp16)[name = tensor<string, []>("cast_262")];
            tensor<fp32, [1, 1, 4, 1]> var_4113 = rsqrt(epsilon = var_4113_epsilon_0, x = var_4112_cast_fp16_to_fp32)[name = tensor<string, []>("op_4113")];
            tensor<string, []> var_4113_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4113_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_4113_to_fp16 = cast(dtype = var_4113_to_fp16_dtype_0, x = var_4113)[name = tensor<string, []>("cast_261")];
            tensor<fp16, [1, 1, 4, 128]> x_945_cast_fp16 = mul(x = x_939_cast_fp16, y = var_4113_to_fp16)[name = tensor<string, []>("x_945_cast_fp16")];
            tensor<fp16, [128]> layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429628800)))];
            tensor<fp16, [1, 1, 4, 128]> var_4115_cast_fp16 = mul(x = layers_23_self_attn_k_norm_weight_to_fp16, y = x_945_cast_fp16)[name = tensor<string, []>("op_4115_cast_fp16")];
            tensor<int32, [4]> k_93_perm_0 = const()[name = tensor<string, []>("k_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(429629120)))];
            tensor<fp16, [1, 1, 512]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = x_929_cast_fp16)[name = tensor<string, []>("linear_163_cast_fp16")];
            tensor<int32, [4]> var_4119 = const()[name = tensor<string, []>("op_4119"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_4120_cast_fp16 = reshape(shape = var_4119, x = linear_163_cast_fp16)[name = tensor<string, []>("op_4120_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = tensor<string, []>("v_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_93_cast_fp16 = transpose(perm = q_93_perm_0, x = var_4099_cast_fp16)[name = tensor<string, []>("transpose_19")];
            tensor<fp16, [1, 12, 1, 128]> var_4124_cast_fp16 = mul(x = q_93_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_4124_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = tensor<string, []>("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = tensor<string, []>("x1_93_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = tensor<string, []>("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_93_cast_fp16)[name = tensor<string, []>("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = tensor<string, []>("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = tensor<string, []>("x2_93_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = tensor<string, []>("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_93_cast_fp16)[name = tensor<string, []>("x2_93_cast_fp16")];
            tensor<fp16, []> const_47_promoted_to_fp16 = const()[name = tensor<string, []>("const_47_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_4127_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_47_promoted_to_fp16)[name = tensor<string, []>("op_4127_cast_fp16")];
            tensor<bool, []> var_4129_interleave_0 = const()[name = tensor<string, []>("op_4129_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_4129_cast_fp16 = concat(axis = var_4058, interleave = var_4129_interleave_0, values = (var_4127_cast_fp16, x1_93_cast_fp16))[name = tensor<string, []>("op_4129_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_4130_cast_fp16 = mul(x = var_4129_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_4130_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_95_cast_fp16 = add(x = var_4124_cast_fp16, y = var_4130_cast_fp16)[name = tensor<string, []>("q_95_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_93_cast_fp16 = transpose(perm = k_93_perm_0, x = var_4115_cast_fp16)[name = tensor<string, []>("transpose_18")];
            tensor<fp16, [1, 4, 1, 128]> var_4132_cast_fp16 = mul(x = k_93_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_4132_cast_fp16")];
            tensor<int32, [4]> x1_95_begin_0 = const()[name = tensor<string, []>("x1_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_95_end_0 = const()[name = tensor<string, []>("x1_95_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_95_end_mask_0 = const()[name = tensor<string, []>("x1_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_95_cast_fp16 = slice_by_index(begin = x1_95_begin_0, end = x1_95_end_0, end_mask = x1_95_end_mask_0, x = k_93_cast_fp16)[name = tensor<string, []>("x1_95_cast_fp16")];
            tensor<int32, [4]> x2_95_begin_0 = const()[name = tensor<string, []>("x2_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_95_end_0 = const()[name = tensor<string, []>("x2_95_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_95_end_mask_0 = const()[name = tensor<string, []>("x2_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_95_cast_fp16 = slice_by_index(begin = x2_95_begin_0, end = x2_95_end_0, end_mask = x2_95_end_mask_0, x = k_93_cast_fp16)[name = tensor<string, []>("x2_95_cast_fp16")];
            tensor<fp16, []> const_48_promoted_to_fp16 = const()[name = tensor<string, []>("const_48_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_4135_cast_fp16 = mul(x = x2_95_cast_fp16, y = const_48_promoted_to_fp16)[name = tensor<string, []>("op_4135_cast_fp16")];
            tensor<bool, []> var_4137_interleave_0 = const()[name = tensor<string, []>("op_4137_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_4137_cast_fp16 = concat(axis = var_4058, interleave = var_4137_interleave_0, values = (var_4135_cast_fp16, x1_95_cast_fp16))[name = tensor<string, []>("op_4137_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_4138_cast_fp16 = mul(x = var_4137_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_4138_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_95_cast_fp16 = add(x = var_4132_cast_fp16, y = var_4138_cast_fp16)[name = tensor<string, []>("k_95_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4141_cast_fp16 = mul(x = k_cache_47_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_4141_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4142_cast_fp16 = mul(x = k_95_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_4142_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_47_cast_fp16 = add(x = var_4141_cast_fp16, y = var_4142_cast_fp16)[name = tensor<string, []>("k_full_47_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4145_cast_fp16 = mul(x = v_cache_47_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_4145_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = var_4120_cast_fp16)[name = tensor<string, []>("transpose_17")];
            tensor<fp16, [1, 4, 2048, 128]> var_4146_cast_fp16 = mul(x = v_47_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_4146_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_47_cast_fp16 = add(x = var_4145_cast_fp16, y = var_4146_cast_fp16)[name = tensor<string, []>("v_full_47_cast_fp16")];
            tensor<int32, [1]> var_4148_axes_0 = const()[name = tensor<string, []>("op_4148_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_4148_cast_fp16 = expand_dims(axes = var_4148_axes_0, x = k_full_47_cast_fp16)[name = tensor<string, []>("op_4148_cast_fp16")];
            tensor<int32, [5]> var_4150_reps_0 = const()[name = tensor<string, []>("op_4150_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_4150_cast_fp16 = tile(reps = var_4150_reps_0, x = var_4148_cast_fp16)[name = tensor<string, []>("op_4150_cast_fp16")];
            tensor<int32, [4]> var_4151 = const()[name = tensor<string, []>("op_4151"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_47_cast_fp16 = reshape(shape = var_4151, x = var_4150_cast_fp16)[name = tensor<string, []>("k_rep_47_cast_fp16")];
            tensor<int32, [1]> var_4153_axes_0 = const()[name = tensor<string, []>("op_4153_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_4153_cast_fp16 = expand_dims(axes = var_4153_axes_0, x = v_full_47_cast_fp16)[name = tensor<string, []>("op_4153_cast_fp16")];
            tensor<int32, [5]> var_4155_reps_0 = const()[name = tensor<string, []>("op_4155_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_4155_cast_fp16 = tile(reps = var_4155_reps_0, x = var_4153_cast_fp16)[name = tensor<string, []>("op_4155_cast_fp16")];
            tensor<int32, [4]> var_4156 = const()[name = tensor<string, []>("op_4156"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_47_cast_fp16 = reshape(shape = var_4156, x = var_4155_cast_fp16)[name = tensor<string, []>("v_rep_47_cast_fp16")];
            tensor<bool, []> var_4159_transpose_x_1 = const()[name = tensor<string, []>("op_4159_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4159_transpose_y_1 = const()[name = tensor<string, []>("op_4159_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_4159_cast_fp16 = matmul(transpose_x = var_4159_transpose_x_1, transpose_y = var_4159_transpose_y_1, x = q_95_cast_fp16, y = k_rep_47_cast_fp16)[name = tensor<string, []>("op_4159_cast_fp16")];
            tensor<fp16, []> var_4160_to_fp16 = const()[name = tensor<string, []>("op_4160_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_93_cast_fp16 = mul(x = var_4159_cast_fp16, y = var_4160_to_fp16)[name = tensor<string, []>("attn_93_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_95_cast_fp16 = add(x = attn_93_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_95_cast_fp16")];
            tensor<string, []> input_95_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_95_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_95_cast_fp16_to_fp32 = cast(dtype = input_95_cast_fp16_to_fp32_dtype_0, x = input_95_cast_fp16)[name = tensor<string, []>("cast_260")];
            tensor<fp32, [1, 12, 1, 2048]> attn_95 = softmax(axis = var_4058, x = input_95_cast_fp16_to_fp32)[name = tensor<string, []>("attn_95")];
            tensor<bool, []> out_47_transpose_x_0 = const()[name = tensor<string, []>("out_47_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_47_transpose_y_0 = const()[name = tensor<string, []>("out_47_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_95_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_95_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_95_to_fp16 = cast(dtype = attn_95_to_fp16_dtype_0, x = attn_95)[name = tensor<string, []>("cast_259")];
            tensor<fp16, [1, 12, 1, 128]> out_47_cast_fp16 = matmul(transpose_x = out_47_transpose_x_0, transpose_y = out_47_transpose_y_0, x = attn_95_to_fp16, y = v_rep_47_cast_fp16)[name = tensor<string, []>("out_47_cast_fp16")];
            tensor<int32, [4]> var_4165_perm_0 = const()[name = tensor<string, []>("op_4165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4167 = const()[name = tensor<string, []>("op_4167"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_4165_cast_fp16 = transpose(perm = var_4165_perm_0, x = out_47_cast_fp16)[name = tensor<string, []>("transpose_16")];
            tensor<fp16, [1, 1, 1536]> x_947_cast_fp16 = reshape(shape = var_4167, x = var_4165_cast_fp16)[name = tensor<string, []>("x_947_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(430153472)))];
            tensor<fp16, [1, 1, 512]> linear_164_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = x_947_cast_fp16)[name = tensor<string, []>("linear_164_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_949_cast_fp16 = add(x = x_921_cast_fp16, y = linear_164_cast_fp16)[name = tensor<string, []>("x_949_cast_fp16")];
            tensor<string, []> x_949_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_949_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4057_promoted_3 = const()[name = tensor<string, []>("op_4057_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_949_cast_fp16_to_fp32 = cast(dtype = x_949_cast_fp16_to_fp32_dtype_0, x = x_949_cast_fp16)[name = tensor<string, []>("cast_258")];
            tensor<fp32, [1, 1, 512]> var_4178 = pow(x = x_949_cast_fp16_to_fp32, y = var_4057_promoted_3)[name = tensor<string, []>("op_4178")];
            tensor<int32, [1]> var_191_axes_0 = const()[name = tensor<string, []>("var_191_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_191_keep_dims_0 = const()[name = tensor<string, []>("var_191_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_191 = reduce_mean(axes = var_191_axes_0, keep_dims = var_191_keep_dims_0, x = var_4178)[name = tensor<string, []>("var_191")];
            tensor<string, []> var_191_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_191_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4182_to_fp16 = const()[name = tensor<string, []>("op_4182_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_191_to_fp16 = cast(dtype = var_191_to_fp16_dtype_0, x = var_191)[name = tensor<string, []>("cast_257")];
            tensor<fp16, [1, 1, 1]> var_4183_cast_fp16 = add(x = var_191_to_fp16, y = var_4182_to_fp16)[name = tensor<string, []>("op_4183_cast_fp16")];
            tensor<string, []> var_4183_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4183_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4184_epsilon_0 = const()[name = tensor<string, []>("op_4184_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4183_cast_fp16_to_fp32 = cast(dtype = var_4183_cast_fp16_to_fp32_dtype_0, x = var_4183_cast_fp16)[name = tensor<string, []>("cast_256")];
            tensor<fp32, [1, 1, 1]> var_4184 = rsqrt(epsilon = var_4184_epsilon_0, x = var_4183_cast_fp16_to_fp32)[name = tensor<string, []>("op_4184")];
            tensor<string, []> var_4184_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4184_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4184_to_fp16 = cast(dtype = var_4184_to_fp16_dtype_0, x = var_4184)[name = tensor<string, []>("cast_255")];
            tensor<fp16, [1, 1, 512]> x_955_cast_fp16 = mul(x = x_949_cast_fp16, y = var_4184_to_fp16)[name = tensor<string, []>("x_955_cast_fp16")];
            tensor<fp16, [512]> layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431726400)))];
            tensor<fp16, [1, 1, 512]> x_957_cast_fp16 = mul(x = layers_23_post_attention_layernorm_weight_to_fp16, y = x_955_cast_fp16)[name = tensor<string, []>("x_957_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_23_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(431727488)))];
            tensor<fp16, [1, 1, 1536]> linear_165_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_23_mlp_gate_proj_weight_to_fp16, x = x_957_cast_fp16)[name = tensor<string, []>("linear_165_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_4195_cast_fp16 = silu(x = linear_165_cast_fp16)[name = tensor<string, []>("op_4195_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_23_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433300416)))];
            tensor<fp16, [1, 1, 1536]> linear_166_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_23_mlp_up_proj_weight_to_fp16, x = x_957_cast_fp16)[name = tensor<string, []>("linear_166_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_959_cast_fp16 = mul(x = var_4195_cast_fp16, y = linear_166_cast_fp16)[name = tensor<string, []>("x_959_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_23_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_23_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(434873344)))];
            tensor<fp16, [1, 1, 512]> linear_167_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_23_mlp_down_proj_weight_to_fp16, x = x_959_cast_fp16)[name = tensor<string, []>("linear_167_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_961_cast_fp16 = add(x = x_949_cast_fp16, y = linear_167_cast_fp16)[name = tensor<string, []>("x_961_cast_fp16")];
            tensor<string, []> x_961_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_961_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_49_begin_0 = const()[name = tensor<string, []>("k_cache_49_begin_0"), val = tensor<int32, [5]>([24, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_49_end_0 = const()[name = tensor<string, []>("k_cache_49_end_0"), val = tensor<int32, [5]>([25, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_49_end_mask_0 = const()[name = tensor<string, []>("k_cache_49_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_49_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_49_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_49_cast_fp16 = slice_by_index(begin = k_cache_49_begin_0, end = k_cache_49_end_0, end_mask = k_cache_49_end_mask_0, squeeze_mask = k_cache_49_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_49_cast_fp16")];
            tensor<int32, [5]> v_cache_49_begin_0 = const()[name = tensor<string, []>("v_cache_49_begin_0"), val = tensor<int32, [5]>([24, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_49_end_0 = const()[name = tensor<string, []>("v_cache_49_end_0"), val = tensor<int32, [5]>([25, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_49_end_mask_0 = const()[name = tensor<string, []>("v_cache_49_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_49_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_49_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_49_cast_fp16 = slice_by_index(begin = v_cache_49_begin_0, end = v_cache_49_end_0, end_mask = v_cache_49_end_mask_0, squeeze_mask = v_cache_49_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_49_cast_fp16")];
            tensor<int32, []> var_4227 = const()[name = tensor<string, []>("op_4227"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_4226_promoted = const()[name = tensor<string, []>("op_4226_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_961_cast_fp16_to_fp32 = cast(dtype = x_961_cast_fp16_to_fp32_dtype_0, x = x_961_cast_fp16)[name = tensor<string, []>("cast_254")];
            tensor<fp32, [1, 1, 512]> var_4236 = pow(x = x_961_cast_fp16_to_fp32, y = var_4226_promoted)[name = tensor<string, []>("op_4236")];
            tensor<int32, [1]> var_193_axes_0 = const()[name = tensor<string, []>("var_193_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_193_keep_dims_0 = const()[name = tensor<string, []>("var_193_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_193 = reduce_mean(axes = var_193_axes_0, keep_dims = var_193_keep_dims_0, x = var_4236)[name = tensor<string, []>("var_193")];
            tensor<string, []> var_193_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_193_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4240_to_fp16 = const()[name = tensor<string, []>("op_4240_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_193_to_fp16 = cast(dtype = var_193_to_fp16_dtype_0, x = var_193)[name = tensor<string, []>("cast_253")];
            tensor<fp16, [1, 1, 1]> var_4241_cast_fp16 = add(x = var_193_to_fp16, y = var_4240_to_fp16)[name = tensor<string, []>("op_4241_cast_fp16")];
            tensor<string, []> var_4241_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4241_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4242_epsilon_0 = const()[name = tensor<string, []>("op_4242_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4241_cast_fp16_to_fp32 = cast(dtype = var_4241_cast_fp16_to_fp32_dtype_0, x = var_4241_cast_fp16)[name = tensor<string, []>("cast_252")];
            tensor<fp32, [1, 1, 1]> var_4242 = rsqrt(epsilon = var_4242_epsilon_0, x = var_4241_cast_fp16_to_fp32)[name = tensor<string, []>("op_4242")];
            tensor<string, []> var_4242_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4242_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4242_to_fp16 = cast(dtype = var_4242_to_fp16_dtype_0, x = var_4242)[name = tensor<string, []>("cast_251")];
            tensor<fp16, [1, 1, 512]> x_967_cast_fp16 = mul(x = x_961_cast_fp16, y = var_4242_to_fp16)[name = tensor<string, []>("x_967_cast_fp16")];
            tensor<fp16, [512]> layers_24_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436446272)))];
            tensor<fp16, [1, 1, 512]> x_969_cast_fp16 = mul(x = layers_24_input_layernorm_weight_to_fp16, y = x_967_cast_fp16)[name = tensor<string, []>("x_969_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_24_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(436447360)))];
            tensor<fp16, [1, 1, 1536]> linear_168_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_24_self_attn_q_proj_weight_to_fp16, x = x_969_cast_fp16)[name = tensor<string, []>("linear_168_cast_fp16")];
            tensor<int32, [4]> var_4256 = const()[name = tensor<string, []>("op_4256"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_971_cast_fp16 = reshape(shape = var_4256, x = linear_168_cast_fp16)[name = tensor<string, []>("x_971_cast_fp16")];
            tensor<string, []> x_971_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_971_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4226_promoted_1 = const()[name = tensor<string, []>("op_4226_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_971_cast_fp16_to_fp32 = cast(dtype = x_971_cast_fp16_to_fp32_dtype_0, x = x_971_cast_fp16)[name = tensor<string, []>("cast_250")];
            tensor<fp32, [1, 1, 12, 128]> var_4260 = pow(x = x_971_cast_fp16_to_fp32, y = var_4226_promoted_1)[name = tensor<string, []>("op_4260")];
            tensor<int32, [1]> var_195_axes_0 = const()[name = tensor<string, []>("var_195_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_195_keep_dims_0 = const()[name = tensor<string, []>("var_195_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_195 = reduce_mean(axes = var_195_axes_0, keep_dims = var_195_keep_dims_0, x = var_4260)[name = tensor<string, []>("var_195")];
            tensor<string, []> var_195_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_195_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4264_to_fp16 = const()[name = tensor<string, []>("op_4264_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_195_to_fp16 = cast(dtype = var_195_to_fp16_dtype_0, x = var_195)[name = tensor<string, []>("cast_249")];
            tensor<fp16, [1, 1, 12, 1]> var_4265_cast_fp16 = add(x = var_195_to_fp16, y = var_4264_to_fp16)[name = tensor<string, []>("op_4265_cast_fp16")];
            tensor<string, []> var_4265_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4265_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4266_epsilon_0 = const()[name = tensor<string, []>("op_4266_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_4265_cast_fp16_to_fp32 = cast(dtype = var_4265_cast_fp16_to_fp32_dtype_0, x = var_4265_cast_fp16)[name = tensor<string, []>("cast_248")];
            tensor<fp32, [1, 1, 12, 1]> var_4266 = rsqrt(epsilon = var_4266_epsilon_0, x = var_4265_cast_fp16_to_fp32)[name = tensor<string, []>("op_4266")];
            tensor<string, []> var_4266_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4266_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_4266_to_fp16 = cast(dtype = var_4266_to_fp16_dtype_0, x = var_4266)[name = tensor<string, []>("cast_247")];
            tensor<fp16, [1, 1, 12, 128]> x_977_cast_fp16 = mul(x = x_971_cast_fp16, y = var_4266_to_fp16)[name = tensor<string, []>("x_977_cast_fp16")];
            tensor<fp16, [128]> layers_24_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438020288)))];
            tensor<fp16, [1, 1, 12, 128]> var_4268_cast_fp16 = mul(x = layers_24_self_attn_q_norm_weight_to_fp16, y = x_977_cast_fp16)[name = tensor<string, []>("op_4268_cast_fp16")];
            tensor<int32, [4]> q_97_perm_0 = const()[name = tensor<string, []>("q_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_24_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438020608)))];
            tensor<fp16, [1, 1, 512]> linear_169_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_24_self_attn_k_proj_weight_to_fp16, x = x_969_cast_fp16)[name = tensor<string, []>("linear_169_cast_fp16")];
            tensor<int32, [4]> var_4272 = const()[name = tensor<string, []>("op_4272"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_979_cast_fp16 = reshape(shape = var_4272, x = linear_169_cast_fp16)[name = tensor<string, []>("x_979_cast_fp16")];
            tensor<string, []> x_979_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_979_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4226_promoted_2 = const()[name = tensor<string, []>("op_4226_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_979_cast_fp16_to_fp32 = cast(dtype = x_979_cast_fp16_to_fp32_dtype_0, x = x_979_cast_fp16)[name = tensor<string, []>("cast_246")];
            tensor<fp32, [1, 1, 4, 128]> var_4276 = pow(x = x_979_cast_fp16_to_fp32, y = var_4226_promoted_2)[name = tensor<string, []>("op_4276")];
            tensor<int32, [1]> var_197_axes_0 = const()[name = tensor<string, []>("var_197_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_197_keep_dims_0 = const()[name = tensor<string, []>("var_197_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_197 = reduce_mean(axes = var_197_axes_0, keep_dims = var_197_keep_dims_0, x = var_4276)[name = tensor<string, []>("var_197")];
            tensor<string, []> var_197_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_197_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4280_to_fp16 = const()[name = tensor<string, []>("op_4280_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_197_to_fp16 = cast(dtype = var_197_to_fp16_dtype_0, x = var_197)[name = tensor<string, []>("cast_245")];
            tensor<fp16, [1, 1, 4, 1]> var_4281_cast_fp16 = add(x = var_197_to_fp16, y = var_4280_to_fp16)[name = tensor<string, []>("op_4281_cast_fp16")];
            tensor<string, []> var_4281_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4281_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4282_epsilon_0 = const()[name = tensor<string, []>("op_4282_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_4281_cast_fp16_to_fp32 = cast(dtype = var_4281_cast_fp16_to_fp32_dtype_0, x = var_4281_cast_fp16)[name = tensor<string, []>("cast_244")];
            tensor<fp32, [1, 1, 4, 1]> var_4282 = rsqrt(epsilon = var_4282_epsilon_0, x = var_4281_cast_fp16_to_fp32)[name = tensor<string, []>("op_4282")];
            tensor<string, []> var_4282_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4282_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_4282_to_fp16 = cast(dtype = var_4282_to_fp16_dtype_0, x = var_4282)[name = tensor<string, []>("cast_243")];
            tensor<fp16, [1, 1, 4, 128]> x_985_cast_fp16 = mul(x = x_979_cast_fp16, y = var_4282_to_fp16)[name = tensor<string, []>("x_985_cast_fp16")];
            tensor<fp16, [128]> layers_24_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438544960)))];
            tensor<fp16, [1, 1, 4, 128]> var_4284_cast_fp16 = mul(x = layers_24_self_attn_k_norm_weight_to_fp16, y = x_985_cast_fp16)[name = tensor<string, []>("op_4284_cast_fp16")];
            tensor<int32, [4]> k_97_perm_0 = const()[name = tensor<string, []>("k_97_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_24_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(438545280)))];
            tensor<fp16, [1, 1, 512]> linear_170_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_24_self_attn_v_proj_weight_to_fp16, x = x_969_cast_fp16)[name = tensor<string, []>("linear_170_cast_fp16")];
            tensor<int32, [4]> var_4288 = const()[name = tensor<string, []>("op_4288"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_4289_cast_fp16 = reshape(shape = var_4288, x = linear_170_cast_fp16)[name = tensor<string, []>("op_4289_cast_fp16")];
            tensor<int32, [4]> v_49_perm_0 = const()[name = tensor<string, []>("v_49_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_97_cast_fp16 = transpose(perm = q_97_perm_0, x = var_4268_cast_fp16)[name = tensor<string, []>("transpose_15")];
            tensor<fp16, [1, 12, 1, 128]> var_4293_cast_fp16 = mul(x = q_97_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_4293_cast_fp16")];
            tensor<int32, [4]> x1_97_begin_0 = const()[name = tensor<string, []>("x1_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_97_end_0 = const()[name = tensor<string, []>("x1_97_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_97_end_mask_0 = const()[name = tensor<string, []>("x1_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_97_cast_fp16 = slice_by_index(begin = x1_97_begin_0, end = x1_97_end_0, end_mask = x1_97_end_mask_0, x = q_97_cast_fp16)[name = tensor<string, []>("x1_97_cast_fp16")];
            tensor<int32, [4]> x2_97_begin_0 = const()[name = tensor<string, []>("x2_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_97_end_0 = const()[name = tensor<string, []>("x2_97_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_97_end_mask_0 = const()[name = tensor<string, []>("x2_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_97_cast_fp16 = slice_by_index(begin = x2_97_begin_0, end = x2_97_end_0, end_mask = x2_97_end_mask_0, x = q_97_cast_fp16)[name = tensor<string, []>("x2_97_cast_fp16")];
            tensor<fp16, []> const_49_promoted_to_fp16 = const()[name = tensor<string, []>("const_49_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_4296_cast_fp16 = mul(x = x2_97_cast_fp16, y = const_49_promoted_to_fp16)[name = tensor<string, []>("op_4296_cast_fp16")];
            tensor<bool, []> var_4298_interleave_0 = const()[name = tensor<string, []>("op_4298_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_4298_cast_fp16 = concat(axis = var_4227, interleave = var_4298_interleave_0, values = (var_4296_cast_fp16, x1_97_cast_fp16))[name = tensor<string, []>("op_4298_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_4299_cast_fp16 = mul(x = var_4298_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_4299_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_99_cast_fp16 = add(x = var_4293_cast_fp16, y = var_4299_cast_fp16)[name = tensor<string, []>("q_99_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_97_cast_fp16 = transpose(perm = k_97_perm_0, x = var_4284_cast_fp16)[name = tensor<string, []>("transpose_14")];
            tensor<fp16, [1, 4, 1, 128]> var_4301_cast_fp16 = mul(x = k_97_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_4301_cast_fp16")];
            tensor<int32, [4]> x1_99_begin_0 = const()[name = tensor<string, []>("x1_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_99_end_0 = const()[name = tensor<string, []>("x1_99_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_99_end_mask_0 = const()[name = tensor<string, []>("x1_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_99_cast_fp16 = slice_by_index(begin = x1_99_begin_0, end = x1_99_end_0, end_mask = x1_99_end_mask_0, x = k_97_cast_fp16)[name = tensor<string, []>("x1_99_cast_fp16")];
            tensor<int32, [4]> x2_99_begin_0 = const()[name = tensor<string, []>("x2_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_99_end_0 = const()[name = tensor<string, []>("x2_99_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_99_end_mask_0 = const()[name = tensor<string, []>("x2_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_99_cast_fp16 = slice_by_index(begin = x2_99_begin_0, end = x2_99_end_0, end_mask = x2_99_end_mask_0, x = k_97_cast_fp16)[name = tensor<string, []>("x2_99_cast_fp16")];
            tensor<fp16, []> const_50_promoted_to_fp16 = const()[name = tensor<string, []>("const_50_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_4304_cast_fp16 = mul(x = x2_99_cast_fp16, y = const_50_promoted_to_fp16)[name = tensor<string, []>("op_4304_cast_fp16")];
            tensor<bool, []> var_4306_interleave_0 = const()[name = tensor<string, []>("op_4306_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_4306_cast_fp16 = concat(axis = var_4227, interleave = var_4306_interleave_0, values = (var_4304_cast_fp16, x1_99_cast_fp16))[name = tensor<string, []>("op_4306_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_4307_cast_fp16 = mul(x = var_4306_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_4307_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_99_cast_fp16 = add(x = var_4301_cast_fp16, y = var_4307_cast_fp16)[name = tensor<string, []>("k_99_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4310_cast_fp16 = mul(x = k_cache_49_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_4310_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4311_cast_fp16 = mul(x = k_99_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_4311_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_49_cast_fp16 = add(x = var_4310_cast_fp16, y = var_4311_cast_fp16)[name = tensor<string, []>("k_full_49_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4314_cast_fp16 = mul(x = v_cache_49_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_4314_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_49_cast_fp16 = transpose(perm = v_49_perm_0, x = var_4289_cast_fp16)[name = tensor<string, []>("transpose_13")];
            tensor<fp16, [1, 4, 2048, 128]> var_4315_cast_fp16 = mul(x = v_49_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_4315_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_49_cast_fp16 = add(x = var_4314_cast_fp16, y = var_4315_cast_fp16)[name = tensor<string, []>("v_full_49_cast_fp16")];
            tensor<int32, [1]> var_4317_axes_0 = const()[name = tensor<string, []>("op_4317_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_4317_cast_fp16 = expand_dims(axes = var_4317_axes_0, x = k_full_49_cast_fp16)[name = tensor<string, []>("op_4317_cast_fp16")];
            tensor<int32, [5]> var_4319_reps_0 = const()[name = tensor<string, []>("op_4319_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_4319_cast_fp16 = tile(reps = var_4319_reps_0, x = var_4317_cast_fp16)[name = tensor<string, []>("op_4319_cast_fp16")];
            tensor<int32, [4]> var_4320 = const()[name = tensor<string, []>("op_4320"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_49_cast_fp16 = reshape(shape = var_4320, x = var_4319_cast_fp16)[name = tensor<string, []>("k_rep_49_cast_fp16")];
            tensor<int32, [1]> var_4322_axes_0 = const()[name = tensor<string, []>("op_4322_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_4322_cast_fp16 = expand_dims(axes = var_4322_axes_0, x = v_full_49_cast_fp16)[name = tensor<string, []>("op_4322_cast_fp16")];
            tensor<int32, [5]> var_4324_reps_0 = const()[name = tensor<string, []>("op_4324_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_4324_cast_fp16 = tile(reps = var_4324_reps_0, x = var_4322_cast_fp16)[name = tensor<string, []>("op_4324_cast_fp16")];
            tensor<int32, [4]> var_4325 = const()[name = tensor<string, []>("op_4325"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_49_cast_fp16 = reshape(shape = var_4325, x = var_4324_cast_fp16)[name = tensor<string, []>("v_rep_49_cast_fp16")];
            tensor<bool, []> var_4328_transpose_x_1 = const()[name = tensor<string, []>("op_4328_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4328_transpose_y_1 = const()[name = tensor<string, []>("op_4328_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_4328_cast_fp16 = matmul(transpose_x = var_4328_transpose_x_1, transpose_y = var_4328_transpose_y_1, x = q_99_cast_fp16, y = k_rep_49_cast_fp16)[name = tensor<string, []>("op_4328_cast_fp16")];
            tensor<fp16, []> var_4329_to_fp16 = const()[name = tensor<string, []>("op_4329_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_97_cast_fp16 = mul(x = var_4328_cast_fp16, y = var_4329_to_fp16)[name = tensor<string, []>("attn_97_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_99_cast_fp16 = add(x = attn_97_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_99_cast_fp16")];
            tensor<string, []> input_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_99_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_99_cast_fp16_to_fp32 = cast(dtype = input_99_cast_fp16_to_fp32_dtype_0, x = input_99_cast_fp16)[name = tensor<string, []>("cast_242")];
            tensor<fp32, [1, 12, 1, 2048]> attn_99 = softmax(axis = var_4227, x = input_99_cast_fp16_to_fp32)[name = tensor<string, []>("attn_99")];
            tensor<bool, []> out_49_transpose_x_0 = const()[name = tensor<string, []>("out_49_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_49_transpose_y_0 = const()[name = tensor<string, []>("out_49_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_99_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_99_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_99_to_fp16 = cast(dtype = attn_99_to_fp16_dtype_0, x = attn_99)[name = tensor<string, []>("cast_241")];
            tensor<fp16, [1, 12, 1, 128]> out_49_cast_fp16 = matmul(transpose_x = out_49_transpose_x_0, transpose_y = out_49_transpose_y_0, x = attn_99_to_fp16, y = v_rep_49_cast_fp16)[name = tensor<string, []>("out_49_cast_fp16")];
            tensor<int32, [4]> var_4334_perm_0 = const()[name = tensor<string, []>("op_4334_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4336 = const()[name = tensor<string, []>("op_4336"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_4334_cast_fp16 = transpose(perm = var_4334_perm_0, x = out_49_cast_fp16)[name = tensor<string, []>("transpose_12")];
            tensor<fp16, [1, 1, 1536]> x_987_cast_fp16 = reshape(shape = var_4336, x = var_4334_cast_fp16)[name = tensor<string, []>("x_987_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_24_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(439069632)))];
            tensor<fp16, [1, 1, 512]> linear_171_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_24_self_attn_o_proj_weight_to_fp16, x = x_987_cast_fp16)[name = tensor<string, []>("linear_171_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_989_cast_fp16 = add(x = x_961_cast_fp16, y = linear_171_cast_fp16)[name = tensor<string, []>("x_989_cast_fp16")];
            tensor<string, []> x_989_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_989_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4226_promoted_3 = const()[name = tensor<string, []>("op_4226_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_989_cast_fp16_to_fp32 = cast(dtype = x_989_cast_fp16_to_fp32_dtype_0, x = x_989_cast_fp16)[name = tensor<string, []>("cast_240")];
            tensor<fp32, [1, 1, 512]> var_4347 = pow(x = x_989_cast_fp16_to_fp32, y = var_4226_promoted_3)[name = tensor<string, []>("op_4347")];
            tensor<int32, [1]> var_199_axes_0 = const()[name = tensor<string, []>("var_199_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_199_keep_dims_0 = const()[name = tensor<string, []>("var_199_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_199 = reduce_mean(axes = var_199_axes_0, keep_dims = var_199_keep_dims_0, x = var_4347)[name = tensor<string, []>("var_199")];
            tensor<string, []> var_199_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_199_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4351_to_fp16 = const()[name = tensor<string, []>("op_4351_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_199_to_fp16 = cast(dtype = var_199_to_fp16_dtype_0, x = var_199)[name = tensor<string, []>("cast_239")];
            tensor<fp16, [1, 1, 1]> var_4352_cast_fp16 = add(x = var_199_to_fp16, y = var_4351_to_fp16)[name = tensor<string, []>("op_4352_cast_fp16")];
            tensor<string, []> var_4352_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4352_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4353_epsilon_0 = const()[name = tensor<string, []>("op_4353_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4352_cast_fp16_to_fp32 = cast(dtype = var_4352_cast_fp16_to_fp32_dtype_0, x = var_4352_cast_fp16)[name = tensor<string, []>("cast_238")];
            tensor<fp32, [1, 1, 1]> var_4353 = rsqrt(epsilon = var_4353_epsilon_0, x = var_4352_cast_fp16_to_fp32)[name = tensor<string, []>("op_4353")];
            tensor<string, []> var_4353_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4353_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4353_to_fp16 = cast(dtype = var_4353_to_fp16_dtype_0, x = var_4353)[name = tensor<string, []>("cast_237")];
            tensor<fp16, [1, 1, 512]> x_995_cast_fp16 = mul(x = x_989_cast_fp16, y = var_4353_to_fp16)[name = tensor<string, []>("x_995_cast_fp16")];
            tensor<fp16, [512]> layers_24_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440642560)))];
            tensor<fp16, [1, 1, 512]> x_997_cast_fp16 = mul(x = layers_24_post_attention_layernorm_weight_to_fp16, y = x_995_cast_fp16)[name = tensor<string, []>("x_997_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_24_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(440643648)))];
            tensor<fp16, [1, 1, 1536]> linear_172_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_24_mlp_gate_proj_weight_to_fp16, x = x_997_cast_fp16)[name = tensor<string, []>("linear_172_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_4364_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor<string, []>("op_4364_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_24_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(442216576)))];
            tensor<fp16, [1, 1, 1536]> linear_173_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_24_mlp_up_proj_weight_to_fp16, x = x_997_cast_fp16)[name = tensor<string, []>("linear_173_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_999_cast_fp16 = mul(x = var_4364_cast_fp16, y = linear_173_cast_fp16)[name = tensor<string, []>("x_999_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_24_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_24_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(443789504)))];
            tensor<fp16, [1, 1, 512]> linear_174_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_24_mlp_down_proj_weight_to_fp16, x = x_999_cast_fp16)[name = tensor<string, []>("linear_174_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_1001_cast_fp16 = add(x = x_989_cast_fp16, y = linear_174_cast_fp16)[name = tensor<string, []>("x_1001_cast_fp16")];
            tensor<string, []> x_1001_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1001_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_51_begin_0 = const()[name = tensor<string, []>("k_cache_51_begin_0"), val = tensor<int32, [5]>([25, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_51_end_0 = const()[name = tensor<string, []>("k_cache_51_end_0"), val = tensor<int32, [5]>([26, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_51_end_mask_0 = const()[name = tensor<string, []>("k_cache_51_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_51_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_51_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_51_cast_fp16 = slice_by_index(begin = k_cache_51_begin_0, end = k_cache_51_end_0, end_mask = k_cache_51_end_mask_0, squeeze_mask = k_cache_51_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_51_cast_fp16")];
            tensor<int32, [5]> v_cache_51_begin_0 = const()[name = tensor<string, []>("v_cache_51_begin_0"), val = tensor<int32, [5]>([25, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_51_end_0 = const()[name = tensor<string, []>("v_cache_51_end_0"), val = tensor<int32, [5]>([26, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_51_end_mask_0 = const()[name = tensor<string, []>("v_cache_51_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_51_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_51_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_51_cast_fp16 = slice_by_index(begin = v_cache_51_begin_0, end = v_cache_51_end_0, end_mask = v_cache_51_end_mask_0, squeeze_mask = v_cache_51_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_51_cast_fp16")];
            tensor<int32, []> var_4396 = const()[name = tensor<string, []>("op_4396"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_4395_promoted = const()[name = tensor<string, []>("op_4395_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_1001_cast_fp16_to_fp32 = cast(dtype = x_1001_cast_fp16_to_fp32_dtype_0, x = x_1001_cast_fp16)[name = tensor<string, []>("cast_236")];
            tensor<fp32, [1, 1, 512]> var_4405 = pow(x = x_1001_cast_fp16_to_fp32, y = var_4395_promoted)[name = tensor<string, []>("op_4405")];
            tensor<int32, [1]> var_201_axes_0 = const()[name = tensor<string, []>("var_201_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_201_keep_dims_0 = const()[name = tensor<string, []>("var_201_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_201 = reduce_mean(axes = var_201_axes_0, keep_dims = var_201_keep_dims_0, x = var_4405)[name = tensor<string, []>("var_201")];
            tensor<string, []> var_201_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_201_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4409_to_fp16 = const()[name = tensor<string, []>("op_4409_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_201_to_fp16 = cast(dtype = var_201_to_fp16_dtype_0, x = var_201)[name = tensor<string, []>("cast_235")];
            tensor<fp16, [1, 1, 1]> var_4410_cast_fp16 = add(x = var_201_to_fp16, y = var_4409_to_fp16)[name = tensor<string, []>("op_4410_cast_fp16")];
            tensor<string, []> var_4410_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4410_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4411_epsilon_0 = const()[name = tensor<string, []>("op_4411_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4410_cast_fp16_to_fp32 = cast(dtype = var_4410_cast_fp16_to_fp32_dtype_0, x = var_4410_cast_fp16)[name = tensor<string, []>("cast_234")];
            tensor<fp32, [1, 1, 1]> var_4411 = rsqrt(epsilon = var_4411_epsilon_0, x = var_4410_cast_fp16_to_fp32)[name = tensor<string, []>("op_4411")];
            tensor<string, []> var_4411_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4411_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4411_to_fp16 = cast(dtype = var_4411_to_fp16_dtype_0, x = var_4411)[name = tensor<string, []>("cast_233")];
            tensor<fp16, [1, 1, 512]> x_1007_cast_fp16 = mul(x = x_1001_cast_fp16, y = var_4411_to_fp16)[name = tensor<string, []>("x_1007_cast_fp16")];
            tensor<fp16, [512]> layers_25_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445362432)))];
            tensor<fp16, [1, 1, 512]> x_1009_cast_fp16 = mul(x = layers_25_input_layernorm_weight_to_fp16, y = x_1007_cast_fp16)[name = tensor<string, []>("x_1009_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_25_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445363520)))];
            tensor<fp16, [1, 1, 1536]> linear_175_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_25_self_attn_q_proj_weight_to_fp16, x = x_1009_cast_fp16)[name = tensor<string, []>("linear_175_cast_fp16")];
            tensor<int32, [4]> var_4425 = const()[name = tensor<string, []>("op_4425"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_1011_cast_fp16 = reshape(shape = var_4425, x = linear_175_cast_fp16)[name = tensor<string, []>("x_1011_cast_fp16")];
            tensor<string, []> x_1011_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1011_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4395_promoted_1 = const()[name = tensor<string, []>("op_4395_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_1011_cast_fp16_to_fp32 = cast(dtype = x_1011_cast_fp16_to_fp32_dtype_0, x = x_1011_cast_fp16)[name = tensor<string, []>("cast_232")];
            tensor<fp32, [1, 1, 12, 128]> var_4429 = pow(x = x_1011_cast_fp16_to_fp32, y = var_4395_promoted_1)[name = tensor<string, []>("op_4429")];
            tensor<int32, [1]> var_203_axes_0 = const()[name = tensor<string, []>("var_203_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_203_keep_dims_0 = const()[name = tensor<string, []>("var_203_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_203 = reduce_mean(axes = var_203_axes_0, keep_dims = var_203_keep_dims_0, x = var_4429)[name = tensor<string, []>("var_203")];
            tensor<string, []> var_203_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_203_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4433_to_fp16 = const()[name = tensor<string, []>("op_4433_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_203_to_fp16 = cast(dtype = var_203_to_fp16_dtype_0, x = var_203)[name = tensor<string, []>("cast_231")];
            tensor<fp16, [1, 1, 12, 1]> var_4434_cast_fp16 = add(x = var_203_to_fp16, y = var_4433_to_fp16)[name = tensor<string, []>("op_4434_cast_fp16")];
            tensor<string, []> var_4434_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4434_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4435_epsilon_0 = const()[name = tensor<string, []>("op_4435_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_4434_cast_fp16_to_fp32 = cast(dtype = var_4434_cast_fp16_to_fp32_dtype_0, x = var_4434_cast_fp16)[name = tensor<string, []>("cast_230")];
            tensor<fp32, [1, 1, 12, 1]> var_4435 = rsqrt(epsilon = var_4435_epsilon_0, x = var_4434_cast_fp16_to_fp32)[name = tensor<string, []>("op_4435")];
            tensor<string, []> var_4435_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4435_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_4435_to_fp16 = cast(dtype = var_4435_to_fp16_dtype_0, x = var_4435)[name = tensor<string, []>("cast_229")];
            tensor<fp16, [1, 1, 12, 128]> x_1017_cast_fp16 = mul(x = x_1011_cast_fp16, y = var_4435_to_fp16)[name = tensor<string, []>("x_1017_cast_fp16")];
            tensor<fp16, [128]> layers_25_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446936448)))];
            tensor<fp16, [1, 1, 12, 128]> var_4437_cast_fp16 = mul(x = layers_25_self_attn_q_norm_weight_to_fp16, y = x_1017_cast_fp16)[name = tensor<string, []>("op_4437_cast_fp16")];
            tensor<int32, [4]> q_101_perm_0 = const()[name = tensor<string, []>("q_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_25_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(446936768)))];
            tensor<fp16, [1, 1, 512]> linear_176_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_25_self_attn_k_proj_weight_to_fp16, x = x_1009_cast_fp16)[name = tensor<string, []>("linear_176_cast_fp16")];
            tensor<int32, [4]> var_4441 = const()[name = tensor<string, []>("op_4441"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_1019_cast_fp16 = reshape(shape = var_4441, x = linear_176_cast_fp16)[name = tensor<string, []>("x_1019_cast_fp16")];
            tensor<string, []> x_1019_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1019_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4395_promoted_2 = const()[name = tensor<string, []>("op_4395_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_1019_cast_fp16_to_fp32 = cast(dtype = x_1019_cast_fp16_to_fp32_dtype_0, x = x_1019_cast_fp16)[name = tensor<string, []>("cast_228")];
            tensor<fp32, [1, 1, 4, 128]> var_4445 = pow(x = x_1019_cast_fp16_to_fp32, y = var_4395_promoted_2)[name = tensor<string, []>("op_4445")];
            tensor<int32, [1]> var_205_axes_0 = const()[name = tensor<string, []>("var_205_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_205_keep_dims_0 = const()[name = tensor<string, []>("var_205_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_205 = reduce_mean(axes = var_205_axes_0, keep_dims = var_205_keep_dims_0, x = var_4445)[name = tensor<string, []>("var_205")];
            tensor<string, []> var_205_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_205_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4449_to_fp16 = const()[name = tensor<string, []>("op_4449_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_205_to_fp16 = cast(dtype = var_205_to_fp16_dtype_0, x = var_205)[name = tensor<string, []>("cast_227")];
            tensor<fp16, [1, 1, 4, 1]> var_4450_cast_fp16 = add(x = var_205_to_fp16, y = var_4449_to_fp16)[name = tensor<string, []>("op_4450_cast_fp16")];
            tensor<string, []> var_4450_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4450_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4451_epsilon_0 = const()[name = tensor<string, []>("op_4451_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_4450_cast_fp16_to_fp32 = cast(dtype = var_4450_cast_fp16_to_fp32_dtype_0, x = var_4450_cast_fp16)[name = tensor<string, []>("cast_226")];
            tensor<fp32, [1, 1, 4, 1]> var_4451 = rsqrt(epsilon = var_4451_epsilon_0, x = var_4450_cast_fp16_to_fp32)[name = tensor<string, []>("op_4451")];
            tensor<string, []> var_4451_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4451_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_4451_to_fp16 = cast(dtype = var_4451_to_fp16_dtype_0, x = var_4451)[name = tensor<string, []>("cast_225")];
            tensor<fp16, [1, 1, 4, 128]> x_1025_cast_fp16 = mul(x = x_1019_cast_fp16, y = var_4451_to_fp16)[name = tensor<string, []>("x_1025_cast_fp16")];
            tensor<fp16, [128]> layers_25_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447461120)))];
            tensor<fp16, [1, 1, 4, 128]> var_4453_cast_fp16 = mul(x = layers_25_self_attn_k_norm_weight_to_fp16, y = x_1025_cast_fp16)[name = tensor<string, []>("op_4453_cast_fp16")];
            tensor<int32, [4]> k_101_perm_0 = const()[name = tensor<string, []>("k_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_25_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447461440)))];
            tensor<fp16, [1, 1, 512]> linear_177_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_25_self_attn_v_proj_weight_to_fp16, x = x_1009_cast_fp16)[name = tensor<string, []>("linear_177_cast_fp16")];
            tensor<int32, [4]> var_4457 = const()[name = tensor<string, []>("op_4457"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_4458_cast_fp16 = reshape(shape = var_4457, x = linear_177_cast_fp16)[name = tensor<string, []>("op_4458_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = tensor<string, []>("v_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_101_cast_fp16 = transpose(perm = q_101_perm_0, x = var_4437_cast_fp16)[name = tensor<string, []>("transpose_11")];
            tensor<fp16, [1, 12, 1, 128]> var_4462_cast_fp16 = mul(x = q_101_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_4462_cast_fp16")];
            tensor<int32, [4]> x1_101_begin_0 = const()[name = tensor<string, []>("x1_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_101_end_0 = const()[name = tensor<string, []>("x1_101_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_101_end_mask_0 = const()[name = tensor<string, []>("x1_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_101_cast_fp16 = slice_by_index(begin = x1_101_begin_0, end = x1_101_end_0, end_mask = x1_101_end_mask_0, x = q_101_cast_fp16)[name = tensor<string, []>("x1_101_cast_fp16")];
            tensor<int32, [4]> x2_101_begin_0 = const()[name = tensor<string, []>("x2_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_101_end_0 = const()[name = tensor<string, []>("x2_101_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_101_end_mask_0 = const()[name = tensor<string, []>("x2_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_101_cast_fp16 = slice_by_index(begin = x2_101_begin_0, end = x2_101_end_0, end_mask = x2_101_end_mask_0, x = q_101_cast_fp16)[name = tensor<string, []>("x2_101_cast_fp16")];
            tensor<fp16, []> const_51_promoted_to_fp16 = const()[name = tensor<string, []>("const_51_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_4465_cast_fp16 = mul(x = x2_101_cast_fp16, y = const_51_promoted_to_fp16)[name = tensor<string, []>("op_4465_cast_fp16")];
            tensor<bool, []> var_4467_interleave_0 = const()[name = tensor<string, []>("op_4467_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_4467_cast_fp16 = concat(axis = var_4396, interleave = var_4467_interleave_0, values = (var_4465_cast_fp16, x1_101_cast_fp16))[name = tensor<string, []>("op_4467_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_4468_cast_fp16 = mul(x = var_4467_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_4468_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_103_cast_fp16 = add(x = var_4462_cast_fp16, y = var_4468_cast_fp16)[name = tensor<string, []>("q_103_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_101_cast_fp16 = transpose(perm = k_101_perm_0, x = var_4453_cast_fp16)[name = tensor<string, []>("transpose_10")];
            tensor<fp16, [1, 4, 1, 128]> var_4470_cast_fp16 = mul(x = k_101_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_4470_cast_fp16")];
            tensor<int32, [4]> x1_103_begin_0 = const()[name = tensor<string, []>("x1_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_103_end_0 = const()[name = tensor<string, []>("x1_103_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_103_end_mask_0 = const()[name = tensor<string, []>("x1_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_103_cast_fp16 = slice_by_index(begin = x1_103_begin_0, end = x1_103_end_0, end_mask = x1_103_end_mask_0, x = k_101_cast_fp16)[name = tensor<string, []>("x1_103_cast_fp16")];
            tensor<int32, [4]> x2_103_begin_0 = const()[name = tensor<string, []>("x2_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_103_end_0 = const()[name = tensor<string, []>("x2_103_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_103_end_mask_0 = const()[name = tensor<string, []>("x2_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_103_cast_fp16 = slice_by_index(begin = x2_103_begin_0, end = x2_103_end_0, end_mask = x2_103_end_mask_0, x = k_101_cast_fp16)[name = tensor<string, []>("x2_103_cast_fp16")];
            tensor<fp16, []> const_52_promoted_to_fp16 = const()[name = tensor<string, []>("const_52_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_4473_cast_fp16 = mul(x = x2_103_cast_fp16, y = const_52_promoted_to_fp16)[name = tensor<string, []>("op_4473_cast_fp16")];
            tensor<bool, []> var_4475_interleave_0 = const()[name = tensor<string, []>("op_4475_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_4475_cast_fp16 = concat(axis = var_4396, interleave = var_4475_interleave_0, values = (var_4473_cast_fp16, x1_103_cast_fp16))[name = tensor<string, []>("op_4475_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_4476_cast_fp16 = mul(x = var_4475_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_4476_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_103_cast_fp16 = add(x = var_4470_cast_fp16, y = var_4476_cast_fp16)[name = tensor<string, []>("k_103_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4479_cast_fp16 = mul(x = k_cache_51_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_4479_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4480_cast_fp16 = mul(x = k_103_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_4480_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_51_cast_fp16 = add(x = var_4479_cast_fp16, y = var_4480_cast_fp16)[name = tensor<string, []>("k_full_51_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4483_cast_fp16 = mul(x = v_cache_51_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_4483_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_4458_cast_fp16)[name = tensor<string, []>("transpose_9")];
            tensor<fp16, [1, 4, 2048, 128]> var_4484_cast_fp16 = mul(x = v_51_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_4484_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_51_cast_fp16 = add(x = var_4483_cast_fp16, y = var_4484_cast_fp16)[name = tensor<string, []>("v_full_51_cast_fp16")];
            tensor<int32, [1]> var_4486_axes_0 = const()[name = tensor<string, []>("op_4486_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_4486_cast_fp16 = expand_dims(axes = var_4486_axes_0, x = k_full_51_cast_fp16)[name = tensor<string, []>("op_4486_cast_fp16")];
            tensor<int32, [5]> var_4488_reps_0 = const()[name = tensor<string, []>("op_4488_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_4488_cast_fp16 = tile(reps = var_4488_reps_0, x = var_4486_cast_fp16)[name = tensor<string, []>("op_4488_cast_fp16")];
            tensor<int32, [4]> var_4489 = const()[name = tensor<string, []>("op_4489"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_51_cast_fp16 = reshape(shape = var_4489, x = var_4488_cast_fp16)[name = tensor<string, []>("k_rep_51_cast_fp16")];
            tensor<int32, [1]> var_4491_axes_0 = const()[name = tensor<string, []>("op_4491_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_4491_cast_fp16 = expand_dims(axes = var_4491_axes_0, x = v_full_51_cast_fp16)[name = tensor<string, []>("op_4491_cast_fp16")];
            tensor<int32, [5]> var_4493_reps_0 = const()[name = tensor<string, []>("op_4493_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_4493_cast_fp16 = tile(reps = var_4493_reps_0, x = var_4491_cast_fp16)[name = tensor<string, []>("op_4493_cast_fp16")];
            tensor<int32, [4]> var_4494 = const()[name = tensor<string, []>("op_4494"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_51_cast_fp16 = reshape(shape = var_4494, x = var_4493_cast_fp16)[name = tensor<string, []>("v_rep_51_cast_fp16")];
            tensor<bool, []> var_4497_transpose_x_1 = const()[name = tensor<string, []>("op_4497_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4497_transpose_y_1 = const()[name = tensor<string, []>("op_4497_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_4497_cast_fp16 = matmul(transpose_x = var_4497_transpose_x_1, transpose_y = var_4497_transpose_y_1, x = q_103_cast_fp16, y = k_rep_51_cast_fp16)[name = tensor<string, []>("op_4497_cast_fp16")];
            tensor<fp16, []> var_4498_to_fp16 = const()[name = tensor<string, []>("op_4498_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_101_cast_fp16 = mul(x = var_4497_cast_fp16, y = var_4498_to_fp16)[name = tensor<string, []>("attn_101_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_103_cast_fp16 = add(x = attn_101_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_103_cast_fp16")];
            tensor<string, []> input_103_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_103_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_103_cast_fp16_to_fp32 = cast(dtype = input_103_cast_fp16_to_fp32_dtype_0, x = input_103_cast_fp16)[name = tensor<string, []>("cast_224")];
            tensor<fp32, [1, 12, 1, 2048]> attn_103 = softmax(axis = var_4396, x = input_103_cast_fp16_to_fp32)[name = tensor<string, []>("attn_103")];
            tensor<bool, []> out_51_transpose_x_0 = const()[name = tensor<string, []>("out_51_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_51_transpose_y_0 = const()[name = tensor<string, []>("out_51_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_103_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_103_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_103_to_fp16 = cast(dtype = attn_103_to_fp16_dtype_0, x = attn_103)[name = tensor<string, []>("cast_223")];
            tensor<fp16, [1, 12, 1, 128]> out_51_cast_fp16 = matmul(transpose_x = out_51_transpose_x_0, transpose_y = out_51_transpose_y_0, x = attn_103_to_fp16, y = v_rep_51_cast_fp16)[name = tensor<string, []>("out_51_cast_fp16")];
            tensor<int32, [4]> var_4503_perm_0 = const()[name = tensor<string, []>("op_4503_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4505 = const()[name = tensor<string, []>("op_4505"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_4503_cast_fp16 = transpose(perm = var_4503_perm_0, x = out_51_cast_fp16)[name = tensor<string, []>("transpose_8")];
            tensor<fp16, [1, 1, 1536]> x_1027_cast_fp16 = reshape(shape = var_4505, x = var_4503_cast_fp16)[name = tensor<string, []>("x_1027_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_25_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(447985792)))];
            tensor<fp16, [1, 1, 512]> linear_178_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_25_self_attn_o_proj_weight_to_fp16, x = x_1027_cast_fp16)[name = tensor<string, []>("linear_178_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_1029_cast_fp16 = add(x = x_1001_cast_fp16, y = linear_178_cast_fp16)[name = tensor<string, []>("x_1029_cast_fp16")];
            tensor<string, []> x_1029_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1029_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4395_promoted_3 = const()[name = tensor<string, []>("op_4395_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_1029_cast_fp16_to_fp32 = cast(dtype = x_1029_cast_fp16_to_fp32_dtype_0, x = x_1029_cast_fp16)[name = tensor<string, []>("cast_222")];
            tensor<fp32, [1, 1, 512]> var_4516 = pow(x = x_1029_cast_fp16_to_fp32, y = var_4395_promoted_3)[name = tensor<string, []>("op_4516")];
            tensor<int32, [1]> var_207_axes_0 = const()[name = tensor<string, []>("var_207_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_207_keep_dims_0 = const()[name = tensor<string, []>("var_207_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_207 = reduce_mean(axes = var_207_axes_0, keep_dims = var_207_keep_dims_0, x = var_4516)[name = tensor<string, []>("var_207")];
            tensor<string, []> var_207_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_207_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4520_to_fp16 = const()[name = tensor<string, []>("op_4520_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_207_to_fp16 = cast(dtype = var_207_to_fp16_dtype_0, x = var_207)[name = tensor<string, []>("cast_221")];
            tensor<fp16, [1, 1, 1]> var_4521_cast_fp16 = add(x = var_207_to_fp16, y = var_4520_to_fp16)[name = tensor<string, []>("op_4521_cast_fp16")];
            tensor<string, []> var_4521_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4521_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4522_epsilon_0 = const()[name = tensor<string, []>("op_4522_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4521_cast_fp16_to_fp32 = cast(dtype = var_4521_cast_fp16_to_fp32_dtype_0, x = var_4521_cast_fp16)[name = tensor<string, []>("cast_220")];
            tensor<fp32, [1, 1, 1]> var_4522 = rsqrt(epsilon = var_4522_epsilon_0, x = var_4521_cast_fp16_to_fp32)[name = tensor<string, []>("op_4522")];
            tensor<string, []> var_4522_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4522_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4522_to_fp16 = cast(dtype = var_4522_to_fp16_dtype_0, x = var_4522)[name = tensor<string, []>("cast_219")];
            tensor<fp16, [1, 1, 512]> x_1035_cast_fp16 = mul(x = x_1029_cast_fp16, y = var_4522_to_fp16)[name = tensor<string, []>("x_1035_cast_fp16")];
            tensor<fp16, [512]> layers_25_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449558720)))];
            tensor<fp16, [1, 1, 512]> x_1037_cast_fp16 = mul(x = layers_25_post_attention_layernorm_weight_to_fp16, y = x_1035_cast_fp16)[name = tensor<string, []>("x_1037_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_25_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(449559808)))];
            tensor<fp16, [1, 1, 1536]> linear_179_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_25_mlp_gate_proj_weight_to_fp16, x = x_1037_cast_fp16)[name = tensor<string, []>("linear_179_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_4533_cast_fp16 = silu(x = linear_179_cast_fp16)[name = tensor<string, []>("op_4533_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_25_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451132736)))];
            tensor<fp16, [1, 1, 1536]> linear_180_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_25_mlp_up_proj_weight_to_fp16, x = x_1037_cast_fp16)[name = tensor<string, []>("linear_180_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_1039_cast_fp16 = mul(x = var_4533_cast_fp16, y = linear_180_cast_fp16)[name = tensor<string, []>("x_1039_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_25_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_25_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452705664)))];
            tensor<fp16, [1, 1, 512]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_25_mlp_down_proj_weight_to_fp16, x = x_1039_cast_fp16)[name = tensor<string, []>("linear_181_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_1041_cast_fp16 = add(x = x_1029_cast_fp16, y = linear_181_cast_fp16)[name = tensor<string, []>("x_1041_cast_fp16")];
            tensor<string, []> x_1041_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1041_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_53_begin_0 = const()[name = tensor<string, []>("k_cache_53_begin_0"), val = tensor<int32, [5]>([26, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_53_end_0 = const()[name = tensor<string, []>("k_cache_53_end_0"), val = tensor<int32, [5]>([27, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_53_end_mask_0 = const()[name = tensor<string, []>("k_cache_53_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_53_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_53_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_53_cast_fp16 = slice_by_index(begin = k_cache_53_begin_0, end = k_cache_53_end_0, end_mask = k_cache_53_end_mask_0, squeeze_mask = k_cache_53_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_53_cast_fp16")];
            tensor<int32, [5]> v_cache_53_begin_0 = const()[name = tensor<string, []>("v_cache_53_begin_0"), val = tensor<int32, [5]>([26, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_53_end_0 = const()[name = tensor<string, []>("v_cache_53_end_0"), val = tensor<int32, [5]>([27, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_53_end_mask_0 = const()[name = tensor<string, []>("v_cache_53_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_53_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_53_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_53_cast_fp16 = slice_by_index(begin = v_cache_53_begin_0, end = v_cache_53_end_0, end_mask = v_cache_53_end_mask_0, squeeze_mask = v_cache_53_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_53_cast_fp16")];
            tensor<int32, []> var_4565 = const()[name = tensor<string, []>("op_4565"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_4564_promoted = const()[name = tensor<string, []>("op_4564_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_1041_cast_fp16_to_fp32 = cast(dtype = x_1041_cast_fp16_to_fp32_dtype_0, x = x_1041_cast_fp16)[name = tensor<string, []>("cast_218")];
            tensor<fp32, [1, 1, 512]> var_4574 = pow(x = x_1041_cast_fp16_to_fp32, y = var_4564_promoted)[name = tensor<string, []>("op_4574")];
            tensor<int32, [1]> var_209_axes_0 = const()[name = tensor<string, []>("var_209_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_209_keep_dims_0 = const()[name = tensor<string, []>("var_209_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_209 = reduce_mean(axes = var_209_axes_0, keep_dims = var_209_keep_dims_0, x = var_4574)[name = tensor<string, []>("var_209")];
            tensor<string, []> var_209_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_209_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4578_to_fp16 = const()[name = tensor<string, []>("op_4578_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_209_to_fp16 = cast(dtype = var_209_to_fp16_dtype_0, x = var_209)[name = tensor<string, []>("cast_217")];
            tensor<fp16, [1, 1, 1]> var_4579_cast_fp16 = add(x = var_209_to_fp16, y = var_4578_to_fp16)[name = tensor<string, []>("op_4579_cast_fp16")];
            tensor<string, []> var_4579_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4579_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4580_epsilon_0 = const()[name = tensor<string, []>("op_4580_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4579_cast_fp16_to_fp32 = cast(dtype = var_4579_cast_fp16_to_fp32_dtype_0, x = var_4579_cast_fp16)[name = tensor<string, []>("cast_216")];
            tensor<fp32, [1, 1, 1]> var_4580 = rsqrt(epsilon = var_4580_epsilon_0, x = var_4579_cast_fp16_to_fp32)[name = tensor<string, []>("op_4580")];
            tensor<string, []> var_4580_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4580_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4580_to_fp16 = cast(dtype = var_4580_to_fp16_dtype_0, x = var_4580)[name = tensor<string, []>("cast_215")];
            tensor<fp16, [1, 1, 512]> x_1047_cast_fp16 = mul(x = x_1041_cast_fp16, y = var_4580_to_fp16)[name = tensor<string, []>("x_1047_cast_fp16")];
            tensor<fp16, [512]> layers_26_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454278592)))];
            tensor<fp16, [1, 1, 512]> x_1049_cast_fp16 = mul(x = layers_26_input_layernorm_weight_to_fp16, y = x_1047_cast_fp16)[name = tensor<string, []>("x_1049_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_26_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454279680)))];
            tensor<fp16, [1, 1, 1536]> linear_182_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_26_self_attn_q_proj_weight_to_fp16, x = x_1049_cast_fp16)[name = tensor<string, []>("linear_182_cast_fp16")];
            tensor<int32, [4]> var_4594 = const()[name = tensor<string, []>("op_4594"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_1051_cast_fp16 = reshape(shape = var_4594, x = linear_182_cast_fp16)[name = tensor<string, []>("x_1051_cast_fp16")];
            tensor<string, []> x_1051_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1051_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4564_promoted_1 = const()[name = tensor<string, []>("op_4564_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_1051_cast_fp16_to_fp32 = cast(dtype = x_1051_cast_fp16_to_fp32_dtype_0, x = x_1051_cast_fp16)[name = tensor<string, []>("cast_214")];
            tensor<fp32, [1, 1, 12, 128]> var_4598 = pow(x = x_1051_cast_fp16_to_fp32, y = var_4564_promoted_1)[name = tensor<string, []>("op_4598")];
            tensor<int32, [1]> var_211_axes_0 = const()[name = tensor<string, []>("var_211_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_211_keep_dims_0 = const()[name = tensor<string, []>("var_211_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_211 = reduce_mean(axes = var_211_axes_0, keep_dims = var_211_keep_dims_0, x = var_4598)[name = tensor<string, []>("var_211")];
            tensor<string, []> var_211_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_211_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4602_to_fp16 = const()[name = tensor<string, []>("op_4602_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_211_to_fp16 = cast(dtype = var_211_to_fp16_dtype_0, x = var_211)[name = tensor<string, []>("cast_213")];
            tensor<fp16, [1, 1, 12, 1]> var_4603_cast_fp16 = add(x = var_211_to_fp16, y = var_4602_to_fp16)[name = tensor<string, []>("op_4603_cast_fp16")];
            tensor<string, []> var_4603_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4603_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4604_epsilon_0 = const()[name = tensor<string, []>("op_4604_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_4603_cast_fp16_to_fp32 = cast(dtype = var_4603_cast_fp16_to_fp32_dtype_0, x = var_4603_cast_fp16)[name = tensor<string, []>("cast_212")];
            tensor<fp32, [1, 1, 12, 1]> var_4604 = rsqrt(epsilon = var_4604_epsilon_0, x = var_4603_cast_fp16_to_fp32)[name = tensor<string, []>("op_4604")];
            tensor<string, []> var_4604_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4604_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_4604_to_fp16 = cast(dtype = var_4604_to_fp16_dtype_0, x = var_4604)[name = tensor<string, []>("cast_211")];
            tensor<fp16, [1, 1, 12, 128]> x_1057_cast_fp16 = mul(x = x_1051_cast_fp16, y = var_4604_to_fp16)[name = tensor<string, []>("x_1057_cast_fp16")];
            tensor<fp16, [128]> layers_26_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(455852608)))];
            tensor<fp16, [1, 1, 12, 128]> var_4606_cast_fp16 = mul(x = layers_26_self_attn_q_norm_weight_to_fp16, y = x_1057_cast_fp16)[name = tensor<string, []>("op_4606_cast_fp16")];
            tensor<int32, [4]> q_105_perm_0 = const()[name = tensor<string, []>("q_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_26_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(455852928)))];
            tensor<fp16, [1, 1, 512]> linear_183_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_26_self_attn_k_proj_weight_to_fp16, x = x_1049_cast_fp16)[name = tensor<string, []>("linear_183_cast_fp16")];
            tensor<int32, [4]> var_4610 = const()[name = tensor<string, []>("op_4610"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_1059_cast_fp16 = reshape(shape = var_4610, x = linear_183_cast_fp16)[name = tensor<string, []>("x_1059_cast_fp16")];
            tensor<string, []> x_1059_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1059_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4564_promoted_2 = const()[name = tensor<string, []>("op_4564_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_1059_cast_fp16_to_fp32 = cast(dtype = x_1059_cast_fp16_to_fp32_dtype_0, x = x_1059_cast_fp16)[name = tensor<string, []>("cast_210")];
            tensor<fp32, [1, 1, 4, 128]> var_4614 = pow(x = x_1059_cast_fp16_to_fp32, y = var_4564_promoted_2)[name = tensor<string, []>("op_4614")];
            tensor<int32, [1]> var_213_axes_0 = const()[name = tensor<string, []>("var_213_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_213_keep_dims_0 = const()[name = tensor<string, []>("var_213_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_213 = reduce_mean(axes = var_213_axes_0, keep_dims = var_213_keep_dims_0, x = var_4614)[name = tensor<string, []>("var_213")];
            tensor<string, []> var_213_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_213_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4618_to_fp16 = const()[name = tensor<string, []>("op_4618_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_213_to_fp16 = cast(dtype = var_213_to_fp16_dtype_0, x = var_213)[name = tensor<string, []>("cast_209")];
            tensor<fp16, [1, 1, 4, 1]> var_4619_cast_fp16 = add(x = var_213_to_fp16, y = var_4618_to_fp16)[name = tensor<string, []>("op_4619_cast_fp16")];
            tensor<string, []> var_4619_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4619_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4620_epsilon_0 = const()[name = tensor<string, []>("op_4620_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_4619_cast_fp16_to_fp32 = cast(dtype = var_4619_cast_fp16_to_fp32_dtype_0, x = var_4619_cast_fp16)[name = tensor<string, []>("cast_208")];
            tensor<fp32, [1, 1, 4, 1]> var_4620 = rsqrt(epsilon = var_4620_epsilon_0, x = var_4619_cast_fp16_to_fp32)[name = tensor<string, []>("op_4620")];
            tensor<string, []> var_4620_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4620_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_4620_to_fp16 = cast(dtype = var_4620_to_fp16_dtype_0, x = var_4620)[name = tensor<string, []>("cast_207")];
            tensor<fp16, [1, 1, 4, 128]> x_1065_cast_fp16 = mul(x = x_1059_cast_fp16, y = var_4620_to_fp16)[name = tensor<string, []>("x_1065_cast_fp16")];
            tensor<fp16, [128]> layers_26_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456377280)))];
            tensor<fp16, [1, 1, 4, 128]> var_4622_cast_fp16 = mul(x = layers_26_self_attn_k_norm_weight_to_fp16, y = x_1065_cast_fp16)[name = tensor<string, []>("op_4622_cast_fp16")];
            tensor<int32, [4]> k_105_perm_0 = const()[name = tensor<string, []>("k_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_26_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456377600)))];
            tensor<fp16, [1, 1, 512]> linear_184_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_26_self_attn_v_proj_weight_to_fp16, x = x_1049_cast_fp16)[name = tensor<string, []>("linear_184_cast_fp16")];
            tensor<int32, [4]> var_4626 = const()[name = tensor<string, []>("op_4626"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_4627_cast_fp16 = reshape(shape = var_4626, x = linear_184_cast_fp16)[name = tensor<string, []>("op_4627_cast_fp16")];
            tensor<int32, [4]> v_53_perm_0 = const()[name = tensor<string, []>("v_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_105_cast_fp16 = transpose(perm = q_105_perm_0, x = var_4606_cast_fp16)[name = tensor<string, []>("transpose_7")];
            tensor<fp16, [1, 12, 1, 128]> var_4631_cast_fp16 = mul(x = q_105_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_4631_cast_fp16")];
            tensor<int32, [4]> x1_105_begin_0 = const()[name = tensor<string, []>("x1_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_105_end_0 = const()[name = tensor<string, []>("x1_105_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_105_end_mask_0 = const()[name = tensor<string, []>("x1_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_105_cast_fp16 = slice_by_index(begin = x1_105_begin_0, end = x1_105_end_0, end_mask = x1_105_end_mask_0, x = q_105_cast_fp16)[name = tensor<string, []>("x1_105_cast_fp16")];
            tensor<int32, [4]> x2_105_begin_0 = const()[name = tensor<string, []>("x2_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_105_end_0 = const()[name = tensor<string, []>("x2_105_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_105_end_mask_0 = const()[name = tensor<string, []>("x2_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_105_cast_fp16 = slice_by_index(begin = x2_105_begin_0, end = x2_105_end_0, end_mask = x2_105_end_mask_0, x = q_105_cast_fp16)[name = tensor<string, []>("x2_105_cast_fp16")];
            tensor<fp16, []> const_53_promoted_to_fp16 = const()[name = tensor<string, []>("const_53_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_4634_cast_fp16 = mul(x = x2_105_cast_fp16, y = const_53_promoted_to_fp16)[name = tensor<string, []>("op_4634_cast_fp16")];
            tensor<bool, []> var_4636_interleave_0 = const()[name = tensor<string, []>("op_4636_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_4636_cast_fp16 = concat(axis = var_4565, interleave = var_4636_interleave_0, values = (var_4634_cast_fp16, x1_105_cast_fp16))[name = tensor<string, []>("op_4636_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_4637_cast_fp16 = mul(x = var_4636_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_4637_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_107_cast_fp16 = add(x = var_4631_cast_fp16, y = var_4637_cast_fp16)[name = tensor<string, []>("q_107_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_105_cast_fp16 = transpose(perm = k_105_perm_0, x = var_4622_cast_fp16)[name = tensor<string, []>("transpose_6")];
            tensor<fp16, [1, 4, 1, 128]> var_4639_cast_fp16 = mul(x = k_105_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_4639_cast_fp16")];
            tensor<int32, [4]> x1_107_begin_0 = const()[name = tensor<string, []>("x1_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_107_end_0 = const()[name = tensor<string, []>("x1_107_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_107_end_mask_0 = const()[name = tensor<string, []>("x1_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_107_cast_fp16 = slice_by_index(begin = x1_107_begin_0, end = x1_107_end_0, end_mask = x1_107_end_mask_0, x = k_105_cast_fp16)[name = tensor<string, []>("x1_107_cast_fp16")];
            tensor<int32, [4]> x2_107_begin_0 = const()[name = tensor<string, []>("x2_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_107_end_0 = const()[name = tensor<string, []>("x2_107_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_107_end_mask_0 = const()[name = tensor<string, []>("x2_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_107_cast_fp16 = slice_by_index(begin = x2_107_begin_0, end = x2_107_end_0, end_mask = x2_107_end_mask_0, x = k_105_cast_fp16)[name = tensor<string, []>("x2_107_cast_fp16")];
            tensor<fp16, []> const_54_promoted_to_fp16 = const()[name = tensor<string, []>("const_54_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_4642_cast_fp16 = mul(x = x2_107_cast_fp16, y = const_54_promoted_to_fp16)[name = tensor<string, []>("op_4642_cast_fp16")];
            tensor<bool, []> var_4644_interleave_0 = const()[name = tensor<string, []>("op_4644_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_4644_cast_fp16 = concat(axis = var_4565, interleave = var_4644_interleave_0, values = (var_4642_cast_fp16, x1_107_cast_fp16))[name = tensor<string, []>("op_4644_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_4645_cast_fp16 = mul(x = var_4644_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_4645_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_107_cast_fp16 = add(x = var_4639_cast_fp16, y = var_4645_cast_fp16)[name = tensor<string, []>("k_107_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4648_cast_fp16 = mul(x = k_cache_53_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_4648_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4649_cast_fp16 = mul(x = k_107_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_4649_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_53_cast_fp16 = add(x = var_4648_cast_fp16, y = var_4649_cast_fp16)[name = tensor<string, []>("k_full_53_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4652_cast_fp16 = mul(x = v_cache_53_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_4652_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_53_cast_fp16 = transpose(perm = v_53_perm_0, x = var_4627_cast_fp16)[name = tensor<string, []>("transpose_5")];
            tensor<fp16, [1, 4, 2048, 128]> var_4653_cast_fp16 = mul(x = v_53_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_4653_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_53_cast_fp16 = add(x = var_4652_cast_fp16, y = var_4653_cast_fp16)[name = tensor<string, []>("v_full_53_cast_fp16")];
            tensor<int32, [1]> var_4655_axes_0 = const()[name = tensor<string, []>("op_4655_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_4655_cast_fp16 = expand_dims(axes = var_4655_axes_0, x = k_full_53_cast_fp16)[name = tensor<string, []>("op_4655_cast_fp16")];
            tensor<int32, [5]> var_4657_reps_0 = const()[name = tensor<string, []>("op_4657_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_4657_cast_fp16 = tile(reps = var_4657_reps_0, x = var_4655_cast_fp16)[name = tensor<string, []>("op_4657_cast_fp16")];
            tensor<int32, [4]> var_4658 = const()[name = tensor<string, []>("op_4658"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_53_cast_fp16 = reshape(shape = var_4658, x = var_4657_cast_fp16)[name = tensor<string, []>("k_rep_53_cast_fp16")];
            tensor<int32, [1]> var_4660_axes_0 = const()[name = tensor<string, []>("op_4660_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_4660_cast_fp16 = expand_dims(axes = var_4660_axes_0, x = v_full_53_cast_fp16)[name = tensor<string, []>("op_4660_cast_fp16")];
            tensor<int32, [5]> var_4662_reps_0 = const()[name = tensor<string, []>("op_4662_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_4662_cast_fp16 = tile(reps = var_4662_reps_0, x = var_4660_cast_fp16)[name = tensor<string, []>("op_4662_cast_fp16")];
            tensor<int32, [4]> var_4663 = const()[name = tensor<string, []>("op_4663"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_53_cast_fp16 = reshape(shape = var_4663, x = var_4662_cast_fp16)[name = tensor<string, []>("v_rep_53_cast_fp16")];
            tensor<bool, []> var_4666_transpose_x_1 = const()[name = tensor<string, []>("op_4666_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4666_transpose_y_1 = const()[name = tensor<string, []>("op_4666_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_4666_cast_fp16 = matmul(transpose_x = var_4666_transpose_x_1, transpose_y = var_4666_transpose_y_1, x = q_107_cast_fp16, y = k_rep_53_cast_fp16)[name = tensor<string, []>("op_4666_cast_fp16")];
            tensor<fp16, []> var_4667_to_fp16 = const()[name = tensor<string, []>("op_4667_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_105_cast_fp16 = mul(x = var_4666_cast_fp16, y = var_4667_to_fp16)[name = tensor<string, []>("attn_105_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_107_cast_fp16 = add(x = attn_105_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_107_cast_fp16")];
            tensor<string, []> input_107_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_107_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_107_cast_fp16_to_fp32 = cast(dtype = input_107_cast_fp16_to_fp32_dtype_0, x = input_107_cast_fp16)[name = tensor<string, []>("cast_206")];
            tensor<fp32, [1, 12, 1, 2048]> attn_107 = softmax(axis = var_4565, x = input_107_cast_fp16_to_fp32)[name = tensor<string, []>("attn_107")];
            tensor<bool, []> out_53_transpose_x_0 = const()[name = tensor<string, []>("out_53_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_53_transpose_y_0 = const()[name = tensor<string, []>("out_53_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_107_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_107_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_107_to_fp16 = cast(dtype = attn_107_to_fp16_dtype_0, x = attn_107)[name = tensor<string, []>("cast_205")];
            tensor<fp16, [1, 12, 1, 128]> out_53_cast_fp16 = matmul(transpose_x = out_53_transpose_x_0, transpose_y = out_53_transpose_y_0, x = attn_107_to_fp16, y = v_rep_53_cast_fp16)[name = tensor<string, []>("out_53_cast_fp16")];
            tensor<int32, [4]> var_4672_perm_0 = const()[name = tensor<string, []>("op_4672_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4674 = const()[name = tensor<string, []>("op_4674"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_4672_cast_fp16 = transpose(perm = var_4672_perm_0, x = out_53_cast_fp16)[name = tensor<string, []>("transpose_4")];
            tensor<fp16, [1, 1, 1536]> x_1067_cast_fp16 = reshape(shape = var_4674, x = var_4672_cast_fp16)[name = tensor<string, []>("x_1067_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_26_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456901952)))];
            tensor<fp16, [1, 1, 512]> linear_185_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_26_self_attn_o_proj_weight_to_fp16, x = x_1067_cast_fp16)[name = tensor<string, []>("linear_185_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_1069_cast_fp16 = add(x = x_1041_cast_fp16, y = linear_185_cast_fp16)[name = tensor<string, []>("x_1069_cast_fp16")];
            tensor<string, []> x_1069_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1069_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4564_promoted_3 = const()[name = tensor<string, []>("op_4564_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_1069_cast_fp16_to_fp32 = cast(dtype = x_1069_cast_fp16_to_fp32_dtype_0, x = x_1069_cast_fp16)[name = tensor<string, []>("cast_204")];
            tensor<fp32, [1, 1, 512]> var_4685 = pow(x = x_1069_cast_fp16_to_fp32, y = var_4564_promoted_3)[name = tensor<string, []>("op_4685")];
            tensor<int32, [1]> var_215_axes_0 = const()[name = tensor<string, []>("var_215_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_215_keep_dims_0 = const()[name = tensor<string, []>("var_215_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_215 = reduce_mean(axes = var_215_axes_0, keep_dims = var_215_keep_dims_0, x = var_4685)[name = tensor<string, []>("var_215")];
            tensor<string, []> var_215_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_215_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4689_to_fp16 = const()[name = tensor<string, []>("op_4689_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_215_to_fp16 = cast(dtype = var_215_to_fp16_dtype_0, x = var_215)[name = tensor<string, []>("cast_203")];
            tensor<fp16, [1, 1, 1]> var_4690_cast_fp16 = add(x = var_215_to_fp16, y = var_4689_to_fp16)[name = tensor<string, []>("op_4690_cast_fp16")];
            tensor<string, []> var_4690_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4690_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4691_epsilon_0 = const()[name = tensor<string, []>("op_4691_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4690_cast_fp16_to_fp32 = cast(dtype = var_4690_cast_fp16_to_fp32_dtype_0, x = var_4690_cast_fp16)[name = tensor<string, []>("cast_202")];
            tensor<fp32, [1, 1, 1]> var_4691 = rsqrt(epsilon = var_4691_epsilon_0, x = var_4690_cast_fp16_to_fp32)[name = tensor<string, []>("op_4691")];
            tensor<string, []> var_4691_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4691_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4691_to_fp16 = cast(dtype = var_4691_to_fp16_dtype_0, x = var_4691)[name = tensor<string, []>("cast_201")];
            tensor<fp16, [1, 1, 512]> x_1075_cast_fp16 = mul(x = x_1069_cast_fp16, y = var_4691_to_fp16)[name = tensor<string, []>("x_1075_cast_fp16")];
            tensor<fp16, [512]> layers_26_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(458474880)))];
            tensor<fp16, [1, 1, 512]> x_1077_cast_fp16 = mul(x = layers_26_post_attention_layernorm_weight_to_fp16, y = x_1075_cast_fp16)[name = tensor<string, []>("x_1077_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_26_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(458475968)))];
            tensor<fp16, [1, 1, 1536]> linear_186_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_26_mlp_gate_proj_weight_to_fp16, x = x_1077_cast_fp16)[name = tensor<string, []>("linear_186_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_4702_cast_fp16 = silu(x = linear_186_cast_fp16)[name = tensor<string, []>("op_4702_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_26_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(460048896)))];
            tensor<fp16, [1, 1, 1536]> linear_187_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_26_mlp_up_proj_weight_to_fp16, x = x_1077_cast_fp16)[name = tensor<string, []>("linear_187_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_1079_cast_fp16 = mul(x = var_4702_cast_fp16, y = linear_187_cast_fp16)[name = tensor<string, []>("x_1079_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_26_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_26_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461621824)))];
            tensor<fp16, [1, 1, 512]> linear_188_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_26_mlp_down_proj_weight_to_fp16, x = x_1079_cast_fp16)[name = tensor<string, []>("linear_188_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_1081_cast_fp16 = add(x = x_1069_cast_fp16, y = linear_188_cast_fp16)[name = tensor<string, []>("x_1081_cast_fp16")];
            tensor<string, []> x_1081_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1081_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, [5]> k_cache_begin_0 = const()[name = tensor<string, []>("k_cache_begin_0"), val = tensor<int32, [5]>([27, 0, 0, 0, 0])];
            tensor<int32, [5]> k_cache_end_0 = const()[name = tensor<string, []>("k_cache_end_0"), val = tensor<int32, [5]>([28, 1, 4, 2048, 128])];
            tensor<bool, [5]> k_cache_end_mask_0 = const()[name = tensor<string, []>("k_cache_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> k_cache_squeeze_mask_0 = const()[name = tensor<string, []>("k_cache_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> k_cache_cast_fp16 = slice_by_index(begin = k_cache_begin_0, end = k_cache_end_0, end_mask = k_cache_end_mask_0, squeeze_mask = k_cache_squeeze_mask_0, x = kv_k_to_fp16)[name = tensor<string, []>("k_cache_cast_fp16")];
            tensor<int32, [5]> v_cache_begin_0 = const()[name = tensor<string, []>("v_cache_begin_0"), val = tensor<int32, [5]>([27, 0, 0, 0, 0])];
            tensor<int32, [5]> v_cache_end_0 = const()[name = tensor<string, []>("v_cache_end_0"), val = tensor<int32, [5]>([28, 1, 4, 2048, 128])];
            tensor<bool, [5]> v_cache_end_mask_0 = const()[name = tensor<string, []>("v_cache_end_mask_0"), val = tensor<bool, [5]>([false, true, true, true, true])];
            tensor<bool, [5]> v_cache_squeeze_mask_0 = const()[name = tensor<string, []>("v_cache_squeeze_mask_0"), val = tensor<bool, [5]>([true, false, false, false, false])];
            tensor<fp16, [1, 4, 2048, 128]> v_cache_cast_fp16 = slice_by_index(begin = v_cache_begin_0, end = v_cache_end_0, end_mask = v_cache_end_mask_0, squeeze_mask = v_cache_squeeze_mask_0, x = kv_v_to_fp16)[name = tensor<string, []>("v_cache_cast_fp16")];
            tensor<int32, []> var_4734 = const()[name = tensor<string, []>("op_4734"), val = tensor<int32, []>(-1)];
            tensor<fp32, []> var_4733_promoted = const()[name = tensor<string, []>("op_4733_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_1081_cast_fp16_to_fp32 = cast(dtype = x_1081_cast_fp16_to_fp32_dtype_0, x = x_1081_cast_fp16)[name = tensor<string, []>("cast_200")];
            tensor<fp32, [1, 1, 512]> var_4743 = pow(x = x_1081_cast_fp16_to_fp32, y = var_4733_promoted)[name = tensor<string, []>("op_4743")];
            tensor<int32, [1]> var_217_axes_0 = const()[name = tensor<string, []>("var_217_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_217_keep_dims_0 = const()[name = tensor<string, []>("var_217_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_217 = reduce_mean(axes = var_217_axes_0, keep_dims = var_217_keep_dims_0, x = var_4743)[name = tensor<string, []>("var_217")];
            tensor<string, []> var_217_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_217_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4747_to_fp16 = const()[name = tensor<string, []>("op_4747_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_217_to_fp16 = cast(dtype = var_217_to_fp16_dtype_0, x = var_217)[name = tensor<string, []>("cast_199")];
            tensor<fp16, [1, 1, 1]> var_4748_cast_fp16 = add(x = var_217_to_fp16, y = var_4747_to_fp16)[name = tensor<string, []>("op_4748_cast_fp16")];
            tensor<string, []> var_4748_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4748_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4749_epsilon_0 = const()[name = tensor<string, []>("op_4749_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4748_cast_fp16_to_fp32 = cast(dtype = var_4748_cast_fp16_to_fp32_dtype_0, x = var_4748_cast_fp16)[name = tensor<string, []>("cast_198")];
            tensor<fp32, [1, 1, 1]> var_4749 = rsqrt(epsilon = var_4749_epsilon_0, x = var_4748_cast_fp16_to_fp32)[name = tensor<string, []>("op_4749")];
            tensor<string, []> var_4749_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4749_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4749_to_fp16 = cast(dtype = var_4749_to_fp16_dtype_0, x = var_4749)[name = tensor<string, []>("cast_197")];
            tensor<fp16, [1, 1, 512]> x_1087_cast_fp16 = mul(x = x_1081_cast_fp16, y = var_4749_to_fp16)[name = tensor<string, []>("x_1087_cast_fp16")];
            tensor<fp16, [512]> layers_27_input_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_input_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(463194752)))];
            tensor<fp16, [1, 1, 512]> x_1089_cast_fp16 = mul(x = layers_27_input_layernorm_weight_to_fp16, y = x_1087_cast_fp16)[name = tensor<string, []>("x_1089_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_27_self_attn_q_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(463195840)))];
            tensor<fp16, [1, 1, 1536]> linear_189_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_27_self_attn_q_proj_weight_to_fp16, x = x_1089_cast_fp16)[name = tensor<string, []>("linear_189_cast_fp16")];
            tensor<int32, [4]> var_4763 = const()[name = tensor<string, []>("op_4763"), val = tensor<int32, [4]>([1, 1, 12, 128])];
            tensor<fp16, [1, 1, 12, 128]> x_1091_cast_fp16 = reshape(shape = var_4763, x = linear_189_cast_fp16)[name = tensor<string, []>("x_1091_cast_fp16")];
            tensor<string, []> x_1091_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1091_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4733_promoted_1 = const()[name = tensor<string, []>("op_4733_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 12, 128]> x_1091_cast_fp16_to_fp32 = cast(dtype = x_1091_cast_fp16_to_fp32_dtype_0, x = x_1091_cast_fp16)[name = tensor<string, []>("cast_196")];
            tensor<fp32, [1, 1, 12, 128]> var_4767 = pow(x = x_1091_cast_fp16_to_fp32, y = var_4733_promoted_1)[name = tensor<string, []>("op_4767")];
            tensor<int32, [1]> var_219_axes_0 = const()[name = tensor<string, []>("var_219_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_219_keep_dims_0 = const()[name = tensor<string, []>("var_219_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 12, 1]> var_219 = reduce_mean(axes = var_219_axes_0, keep_dims = var_219_keep_dims_0, x = var_4767)[name = tensor<string, []>("var_219")];
            tensor<string, []> var_219_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_219_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4771_to_fp16 = const()[name = tensor<string, []>("op_4771_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 12, 1]> var_219_to_fp16 = cast(dtype = var_219_to_fp16_dtype_0, x = var_219)[name = tensor<string, []>("cast_195")];
            tensor<fp16, [1, 1, 12, 1]> var_4772_cast_fp16 = add(x = var_219_to_fp16, y = var_4771_to_fp16)[name = tensor<string, []>("op_4772_cast_fp16")];
            tensor<string, []> var_4772_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4772_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4773_epsilon_0 = const()[name = tensor<string, []>("op_4773_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 12, 1]> var_4772_cast_fp16_to_fp32 = cast(dtype = var_4772_cast_fp16_to_fp32_dtype_0, x = var_4772_cast_fp16)[name = tensor<string, []>("cast_194")];
            tensor<fp32, [1, 1, 12, 1]> var_4773 = rsqrt(epsilon = var_4773_epsilon_0, x = var_4772_cast_fp16_to_fp32)[name = tensor<string, []>("op_4773")];
            tensor<string, []> var_4773_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4773_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 12, 1]> var_4773_to_fp16 = cast(dtype = var_4773_to_fp16_dtype_0, x = var_4773)[name = tensor<string, []>("cast_193")];
            tensor<fp16, [1, 1, 12, 128]> x_1097_cast_fp16 = mul(x = x_1091_cast_fp16, y = var_4773_to_fp16)[name = tensor<string, []>("x_1097_cast_fp16")];
            tensor<fp16, [128]> layers_27_self_attn_q_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464768768)))];
            tensor<fp16, [1, 1, 12, 128]> var_4775_cast_fp16 = mul(x = layers_27_self_attn_q_norm_weight_to_fp16, y = x_1097_cast_fp16)[name = tensor<string, []>("op_4775_cast_fp16")];
            tensor<int32, [4]> q_109_perm_0 = const()[name = tensor<string, []>("q_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_27_self_attn_k_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(464769088)))];
            tensor<fp16, [1, 1, 512]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_27_self_attn_k_proj_weight_to_fp16, x = x_1089_cast_fp16)[name = tensor<string, []>("linear_190_cast_fp16")];
            tensor<int32, [4]> var_4779 = const()[name = tensor<string, []>("op_4779"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> x_1099_cast_fp16 = reshape(shape = var_4779, x = linear_190_cast_fp16)[name = tensor<string, []>("x_1099_cast_fp16")];
            tensor<string, []> x_1099_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1099_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4733_promoted_2 = const()[name = tensor<string, []>("op_4733_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 4, 128]> x_1099_cast_fp16_to_fp32 = cast(dtype = x_1099_cast_fp16_to_fp32_dtype_0, x = x_1099_cast_fp16)[name = tensor<string, []>("cast_192")];
            tensor<fp32, [1, 1, 4, 128]> var_4783 = pow(x = x_1099_cast_fp16_to_fp32, y = var_4733_promoted_2)[name = tensor<string, []>("op_4783")];
            tensor<int32, [1]> var_221_axes_0 = const()[name = tensor<string, []>("var_221_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_221_keep_dims_0 = const()[name = tensor<string, []>("var_221_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 4, 1]> var_221 = reduce_mean(axes = var_221_axes_0, keep_dims = var_221_keep_dims_0, x = var_4783)[name = tensor<string, []>("var_221")];
            tensor<string, []> var_221_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_221_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4787_to_fp16 = const()[name = tensor<string, []>("op_4787_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 4, 1]> var_221_to_fp16 = cast(dtype = var_221_to_fp16_dtype_0, x = var_221)[name = tensor<string, []>("cast_191")];
            tensor<fp16, [1, 1, 4, 1]> var_4788_cast_fp16 = add(x = var_221_to_fp16, y = var_4787_to_fp16)[name = tensor<string, []>("op_4788_cast_fp16")];
            tensor<string, []> var_4788_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4788_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4789_epsilon_0 = const()[name = tensor<string, []>("op_4789_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 4, 1]> var_4788_cast_fp16_to_fp32 = cast(dtype = var_4788_cast_fp16_to_fp32_dtype_0, x = var_4788_cast_fp16)[name = tensor<string, []>("cast_190")];
            tensor<fp32, [1, 1, 4, 1]> var_4789 = rsqrt(epsilon = var_4789_epsilon_0, x = var_4788_cast_fp16_to_fp32)[name = tensor<string, []>("op_4789")];
            tensor<string, []> var_4789_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4789_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 4, 1]> var_4789_to_fp16 = cast(dtype = var_4789_to_fp16_dtype_0, x = var_4789)[name = tensor<string, []>("cast_189")];
            tensor<fp16, [1, 1, 4, 128]> x_1105_cast_fp16 = mul(x = x_1099_cast_fp16, y = var_4789_to_fp16)[name = tensor<string, []>("x_1105_cast_fp16")];
            tensor<fp16, [128]> layers_27_self_attn_k_norm_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465293440)))];
            tensor<fp16, [1, 1, 4, 128]> var_4791_cast_fp16 = mul(x = layers_27_self_attn_k_norm_weight_to_fp16, y = x_1105_cast_fp16)[name = tensor<string, []>("op_4791_cast_fp16")];
            tensor<int32, [4]> k_109_perm_0 = const()[name = tensor<string, []>("k_109_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [512, 512]> layers_27_self_attn_v_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [512, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465293760)))];
            tensor<fp16, [1, 1, 512]> linear_191_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_27_self_attn_v_proj_weight_to_fp16, x = x_1089_cast_fp16)[name = tensor<string, []>("linear_191_cast_fp16")];
            tensor<int32, [4]> var_4795 = const()[name = tensor<string, []>("op_4795"), val = tensor<int32, [4]>([1, 1, 4, 128])];
            tensor<fp16, [1, 1, 4, 128]> var_4796_cast_fp16 = reshape(shape = var_4795, x = linear_191_cast_fp16)[name = tensor<string, []>("op_4796_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = tensor<string, []>("v_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<fp16, [1, 12, 1, 128]> q_109_cast_fp16 = transpose(perm = q_109_perm_0, x = var_4775_cast_fp16)[name = tensor<string, []>("transpose_3")];
            tensor<fp16, [1, 12, 1, 128]> var_4800_cast_fp16 = mul(x = q_109_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_4800_cast_fp16")];
            tensor<int32, [4]> x1_109_begin_0 = const()[name = tensor<string, []>("x1_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_109_end_0 = const()[name = tensor<string, []>("x1_109_end_0"), val = tensor<int32, [4]>([1, 12, 1, 64])];
            tensor<bool, [4]> x1_109_end_mask_0 = const()[name = tensor<string, []>("x1_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 12, 1, 64]> x1_109_cast_fp16 = slice_by_index(begin = x1_109_begin_0, end = x1_109_end_0, end_mask = x1_109_end_mask_0, x = q_109_cast_fp16)[name = tensor<string, []>("x1_109_cast_fp16")];
            tensor<int32, [4]> x2_109_begin_0 = const()[name = tensor<string, []>("x2_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_109_end_0 = const()[name = tensor<string, []>("x2_109_end_0"), val = tensor<int32, [4]>([1, 12, 1, 128])];
            tensor<bool, [4]> x2_109_end_mask_0 = const()[name = tensor<string, []>("x2_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 12, 1, 64]> x2_109_cast_fp16 = slice_by_index(begin = x2_109_begin_0, end = x2_109_end_0, end_mask = x2_109_end_mask_0, x = q_109_cast_fp16)[name = tensor<string, []>("x2_109_cast_fp16")];
            tensor<fp16, []> const_55_promoted_to_fp16 = const()[name = tensor<string, []>("const_55_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 12, 1, 64]> var_4803_cast_fp16 = mul(x = x2_109_cast_fp16, y = const_55_promoted_to_fp16)[name = tensor<string, []>("op_4803_cast_fp16")];
            tensor<bool, []> var_4805_interleave_0 = const()[name = tensor<string, []>("op_4805_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 12, 1, 128]> var_4805_cast_fp16 = concat(axis = var_4734, interleave = var_4805_interleave_0, values = (var_4803_cast_fp16, x1_109_cast_fp16))[name = tensor<string, []>("op_4805_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> var_4806_cast_fp16 = mul(x = var_4805_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_4806_cast_fp16")];
            tensor<fp16, [1, 12, 1, 128]> q_cast_fp16 = add(x = var_4800_cast_fp16, y = var_4806_cast_fp16)[name = tensor<string, []>("q_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_109_cast_fp16 = transpose(perm = k_109_perm_0, x = var_4791_cast_fp16)[name = tensor<string, []>("transpose_2")];
            tensor<fp16, [1, 4, 1, 128]> var_4808_cast_fp16 = mul(x = k_109_cast_fp16, y = cos_3_cast_fp16)[name = tensor<string, []>("op_4808_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = tensor<string, []>("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = tensor<string, []>("x1_end_0"), val = tensor<int32, [4]>([1, 4, 1, 64])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = tensor<string, []>("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 4, 1, 64]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_109_cast_fp16)[name = tensor<string, []>("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = tensor<string, []>("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> x2_end_0 = const()[name = tensor<string, []>("x2_end_0"), val = tensor<int32, [4]>([1, 4, 1, 128])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = tensor<string, []>("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 4, 1, 64]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_109_cast_fp16)[name = tensor<string, []>("x2_cast_fp16")];
            tensor<fp16, []> const_56_promoted_to_fp16 = const()[name = tensor<string, []>("const_56_promoted_to_fp16"), val = tensor<fp16, []>(-0x1p+0)];
            tensor<fp16, [1, 4, 1, 64]> var_4811_cast_fp16 = mul(x = x2_cast_fp16, y = const_56_promoted_to_fp16)[name = tensor<string, []>("op_4811_cast_fp16")];
            tensor<bool, []> var_4813_interleave_0 = const()[name = tensor<string, []>("op_4813_interleave_0"), val = tensor<bool, []>(false)];
            tensor<fp16, [1, 4, 1, 128]> var_4813_cast_fp16 = concat(axis = var_4734, interleave = var_4813_interleave_0, values = (var_4811_cast_fp16, x1_cast_fp16))[name = tensor<string, []>("op_4813_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> var_4814_cast_fp16 = mul(x = var_4813_cast_fp16, y = sin_3_cast_fp16)[name = tensor<string, []>("op_4814_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> k_cast_fp16 = add(x = var_4808_cast_fp16, y = var_4814_cast_fp16)[name = tensor<string, []>("k_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4817_cast_fp16 = mul(x = k_cache_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_4817_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4818_cast_fp16 = mul(x = k_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_4818_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> k_full_cast_fp16 = add(x = var_4817_cast_fp16, y = var_4818_cast_fp16)[name = tensor<string, []>("k_full_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> var_4821_cast_fp16 = mul(x = v_cache_cast_fp16, y = var_253_cast_fp16)[name = tensor<string, []>("op_4821_cast_fp16")];
            tensor<fp16, [1, 4, 1, 128]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_4796_cast_fp16)[name = tensor<string, []>("transpose_1")];
            tensor<fp16, [1, 4, 2048, 128]> var_4822_cast_fp16 = mul(x = v_cast_fp16, y = var_107_to_fp16)[name = tensor<string, []>("op_4822_cast_fp16")];
            tensor<fp16, [1, 4, 2048, 128]> v_full_cast_fp16 = add(x = var_4821_cast_fp16, y = var_4822_cast_fp16)[name = tensor<string, []>("v_full_cast_fp16")];
            tensor<int32, [1]> var_4824_axes_0 = const()[name = tensor<string, []>("op_4824_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_4824_cast_fp16 = expand_dims(axes = var_4824_axes_0, x = k_full_cast_fp16)[name = tensor<string, []>("op_4824_cast_fp16")];
            tensor<int32, [5]> var_4826_reps_0 = const()[name = tensor<string, []>("op_4826_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_4826_cast_fp16 = tile(reps = var_4826_reps_0, x = var_4824_cast_fp16)[name = tensor<string, []>("op_4826_cast_fp16")];
            tensor<int32, [4]> var_4827 = const()[name = tensor<string, []>("op_4827"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> k_rep_cast_fp16 = reshape(shape = var_4827, x = var_4826_cast_fp16)[name = tensor<string, []>("k_rep_cast_fp16")];
            tensor<int32, [1]> var_4829_axes_0 = const()[name = tensor<string, []>("op_4829_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 4, 1, 2048, 128]> var_4829_cast_fp16 = expand_dims(axes = var_4829_axes_0, x = v_full_cast_fp16)[name = tensor<string, []>("op_4829_cast_fp16")];
            tensor<int32, [5]> var_4831_reps_0 = const()[name = tensor<string, []>("op_4831_reps_0"), val = tensor<int32, [5]>([1, 1, 3, 1, 1])];
            tensor<fp16, [1, 4, 3, 2048, 128]> var_4831_cast_fp16 = tile(reps = var_4831_reps_0, x = var_4829_cast_fp16)[name = tensor<string, []>("op_4831_cast_fp16")];
            tensor<int32, [4]> var_4832 = const()[name = tensor<string, []>("op_4832"), val = tensor<int32, [4]>([1, 12, 2048, 128])];
            tensor<fp16, [1, 12, 2048, 128]> v_rep_cast_fp16 = reshape(shape = var_4832, x = var_4831_cast_fp16)[name = tensor<string, []>("v_rep_cast_fp16")];
            tensor<bool, []> var_4835_transpose_x_1 = const()[name = tensor<string, []>("op_4835_transpose_x_1"), val = tensor<bool, []>(false)];
            tensor<bool, []> var_4835_transpose_y_1 = const()[name = tensor<string, []>("op_4835_transpose_y_1"), val = tensor<bool, []>(true)];
            tensor<fp16, [1, 12, 1, 2048]> var_4835_cast_fp16 = matmul(transpose_x = var_4835_transpose_x_1, transpose_y = var_4835_transpose_y_1, x = q_cast_fp16, y = k_rep_cast_fp16)[name = tensor<string, []>("op_4835_cast_fp16")];
            tensor<fp16, []> var_4836_to_fp16 = const()[name = tensor<string, []>("op_4836_to_fp16"), val = tensor<fp16, []>(0x1.6ap-4)];
            tensor<fp16, [1, 12, 1, 2048]> attn_109_cast_fp16 = mul(x = var_4835_cast_fp16, y = var_4836_to_fp16)[name = tensor<string, []>("attn_109_cast_fp16")];
            tensor<fp16, [1, 12, 1, 2048]> input_111_cast_fp16 = add(x = attn_109_cast_fp16, y = attn_mask_cast_fp16)[name = tensor<string, []>("input_111_cast_fp16")];
            tensor<string, []> input_111_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("input_111_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [1, 12, 1, 2048]> input_111_cast_fp16_to_fp32 = cast(dtype = input_111_cast_fp16_to_fp32_dtype_0, x = input_111_cast_fp16)[name = tensor<string, []>("cast_188")];
            tensor<fp32, [1, 12, 1, 2048]> attn = softmax(axis = var_4734, x = input_111_cast_fp16_to_fp32)[name = tensor<string, []>("attn")];
            tensor<bool, []> out_transpose_x_0 = const()[name = tensor<string, []>("out_transpose_x_0"), val = tensor<bool, []>(false)];
            tensor<bool, []> out_transpose_y_0 = const()[name = tensor<string, []>("out_transpose_y_0"), val = tensor<bool, []>(false)];
            tensor<string, []> attn_to_fp16_dtype_0 = const()[name = tensor<string, []>("attn_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 12, 1, 2048]> attn_to_fp16 = cast(dtype = attn_to_fp16_dtype_0, x = attn)[name = tensor<string, []>("cast_187")];
            tensor<fp16, [1, 12, 1, 128]> out_cast_fp16 = matmul(transpose_x = out_transpose_x_0, transpose_y = out_transpose_y_0, x = attn_to_fp16, y = v_rep_cast_fp16)[name = tensor<string, []>("out_cast_fp16")];
            tensor<int32, [4]> var_4841_perm_0 = const()[name = tensor<string, []>("op_4841_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4843 = const()[name = tensor<string, []>("op_4843"), val = tensor<int32, [3]>([1, 1, 1536])];
            tensor<fp16, [1, 1, 12, 128]> var_4841_cast_fp16 = transpose(perm = var_4841_perm_0, x = out_cast_fp16)[name = tensor<string, []>("transpose_0")];
            tensor<fp16, [1, 1, 1536]> x_1107_cast_fp16 = reshape(shape = var_4843, x = var_4841_cast_fp16)[name = tensor<string, []>("x_1107_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_27_self_attn_o_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465818112)))];
            tensor<fp16, [1, 1, 512]> linear_192_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_27_self_attn_o_proj_weight_to_fp16, x = x_1107_cast_fp16)[name = tensor<string, []>("linear_192_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_1109_cast_fp16 = add(x = x_1081_cast_fp16, y = linear_192_cast_fp16)[name = tensor<string, []>("x_1109_cast_fp16")];
            tensor<string, []> x_1109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1109_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4733_promoted_3 = const()[name = tensor<string, []>("op_4733_promoted_3"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_1109_cast_fp16_to_fp32 = cast(dtype = x_1109_cast_fp16_to_fp32_dtype_0, x = x_1109_cast_fp16)[name = tensor<string, []>("cast_186")];
            tensor<fp32, [1, 1, 512]> var_4854 = pow(x = x_1109_cast_fp16_to_fp32, y = var_4733_promoted_3)[name = tensor<string, []>("op_4854")];
            tensor<int32, [1]> var_223_axes_0 = const()[name = tensor<string, []>("var_223_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_223_keep_dims_0 = const()[name = tensor<string, []>("var_223_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var_223 = reduce_mean(axes = var_223_axes_0, keep_dims = var_223_keep_dims_0, x = var_4854)[name = tensor<string, []>("var_223")];
            tensor<string, []> var_223_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_223_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4858_to_fp16 = const()[name = tensor<string, []>("op_4858_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_223_to_fp16 = cast(dtype = var_223_to_fp16_dtype_0, x = var_223)[name = tensor<string, []>("cast_185")];
            tensor<fp16, [1, 1, 1]> var_4859_cast_fp16 = add(x = var_223_to_fp16, y = var_4858_to_fp16)[name = tensor<string, []>("op_4859_cast_fp16")];
            tensor<string, []> var_4859_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4859_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4860_epsilon_0 = const()[name = tensor<string, []>("op_4860_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4859_cast_fp16_to_fp32 = cast(dtype = var_4859_cast_fp16_to_fp32_dtype_0, x = var_4859_cast_fp16)[name = tensor<string, []>("cast_184")];
            tensor<fp32, [1, 1, 1]> var_4860 = rsqrt(epsilon = var_4860_epsilon_0, x = var_4859_cast_fp16_to_fp32)[name = tensor<string, []>("op_4860")];
            tensor<string, []> var_4860_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4860_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4860_to_fp16 = cast(dtype = var_4860_to_fp16_dtype_0, x = var_4860)[name = tensor<string, []>("cast_183")];
            tensor<fp16, [1, 1, 512]> x_1115_cast_fp16 = mul(x = x_1109_cast_fp16, y = var_4860_to_fp16)[name = tensor<string, []>("x_1115_cast_fp16")];
            tensor<fp16, [512]> layers_27_post_attention_layernorm_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(467391040)))];
            tensor<fp16, [1, 1, 512]> x_1117_cast_fp16 = mul(x = layers_27_post_attention_layernorm_weight_to_fp16, y = x_1115_cast_fp16)[name = tensor<string, []>("x_1117_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_27_mlp_gate_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(467392128)))];
            tensor<fp16, [1, 1, 1536]> linear_193_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_27_mlp_gate_proj_weight_to_fp16, x = x_1117_cast_fp16)[name = tensor<string, []>("linear_193_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> var_4871_cast_fp16 = silu(x = linear_193_cast_fp16)[name = tensor<string, []>("op_4871_cast_fp16")];
            tensor<fp16, [1536, 512]> layers_27_mlp_up_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [1536, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(468965056)))];
            tensor<fp16, [1, 1, 1536]> linear_194_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_27_mlp_up_proj_weight_to_fp16, x = x_1117_cast_fp16)[name = tensor<string, []>("linear_194_cast_fp16")];
            tensor<fp16, [1, 1, 1536]> x_1119_cast_fp16 = mul(x = var_4871_cast_fp16, y = linear_194_cast_fp16)[name = tensor<string, []>("x_1119_cast_fp16")];
            tensor<fp16, [512, 1536]> layers_27_mlp_down_proj_weight_to_fp16 = const()[name = tensor<string, []>("layers_27_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [512, 1536]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(470537984)))];
            tensor<fp16, [1, 1, 512]> linear_195_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = layers_27_mlp_down_proj_weight_to_fp16, x = x_1119_cast_fp16)[name = tensor<string, []>("linear_195_cast_fp16")];
            tensor<fp16, [1, 1, 512]> x_1121_cast_fp16 = add(x = x_1109_cast_fp16, y = linear_195_cast_fp16)[name = tensor<string, []>("x_1121_cast_fp16")];
            tensor<string, []> x_1121_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("x_1121_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4885_promoted = const()[name = tensor<string, []>("op_4885_promoted"), val = tensor<fp32, []>(0x1p+1)];
            tensor<fp32, [1, 1, 512]> x_1121_cast_fp16_to_fp32 = cast(dtype = x_1121_cast_fp16_to_fp32_dtype_0, x = x_1121_cast_fp16)[name = tensor<string, []>("cast_182")];
            tensor<fp32, [1, 1, 512]> var_4891 = pow(x = x_1121_cast_fp16_to_fp32, y = var_4885_promoted)[name = tensor<string, []>("op_4891")];
            tensor<int32, [1]> var_axes_0 = const()[name = tensor<string, []>("var_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<bool, []> var_keep_dims_0 = const()[name = tensor<string, []>("var_keep_dims_0"), val = tensor<bool, []>(true)];
            tensor<fp32, [1, 1, 1]> var = reduce_mean(axes = var_axes_0, keep_dims = var_keep_dims_0, x = var_4891)[name = tensor<string, []>("var")];
            tensor<string, []> var_to_fp16_dtype_0 = const()[name = tensor<string, []>("var_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, []> var_4895_to_fp16 = const()[name = tensor<string, []>("op_4895_to_fp16"), val = tensor<fp16, []>(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_to_fp16 = cast(dtype = var_to_fp16_dtype_0, x = var)[name = tensor<string, []>("cast_181")];
            tensor<fp16, [1, 1, 1]> var_4896_cast_fp16 = add(x = var_to_fp16, y = var_4895_to_fp16)[name = tensor<string, []>("op_4896_cast_fp16")];
            tensor<string, []> var_4896_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4896_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, []> var_4897_epsilon_0 = const()[name = tensor<string, []>("op_4897_epsilon_0"), val = tensor<fp32, []>(0x1.197998p-40)];
            tensor<fp32, [1, 1, 1]> var_4896_cast_fp16_to_fp32 = cast(dtype = var_4896_cast_fp16_to_fp32_dtype_0, x = var_4896_cast_fp16)[name = tensor<string, []>("cast_180")];
            tensor<fp32, [1, 1, 1]> var_4897 = rsqrt(epsilon = var_4897_epsilon_0, x = var_4896_cast_fp16_to_fp32)[name = tensor<string, []>("op_4897")];
            tensor<string, []> var_4897_to_fp16_dtype_0 = const()[name = tensor<string, []>("op_4897_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
            tensor<fp16, [1, 1, 1]> var_4897_to_fp16 = cast(dtype = var_4897_to_fp16_dtype_0, x = var_4897)[name = tensor<string, []>("cast_179")];
            tensor<fp16, [1, 1, 512]> x_cast_fp16 = mul(x = x_1121_cast_fp16, y = var_4897_to_fp16)[name = tensor<string, []>("x_cast_fp16")];
            tensor<fp16, [512]> norm_weight_to_fp16 = const()[name = tensor<string, []>("norm_weight_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(472110912)))];
            tensor<fp16, [1, 1, 512]> hidden_cast_fp16 = mul(x = norm_weight_to_fp16, y = x_cast_fp16)[name = tensor<string, []>("hidden_cast_fp16")];
            tensor<fp16, [217232]> linear_196_bias_0_to_fp16 = const()[name = tensor<string, []>("linear_196_bias_0_to_fp16"), val = tensor<fp16, [217232]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(472112000)))];
            tensor<fp16, [1, 1, 217232]> linear_196_cast_fp16 = linear(bias = linear_196_bias_0_to_fp16, weight = tok_weight_to_fp16, x = hidden_cast_fp16)[name = tensor<string, []>("linear_196_cast_fp16")];
            tensor<int32, [1]> var_4903_axes_0 = const()[name = tensor<string, []>("op_4903_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 217232]> var_4903_cast_fp16 = squeeze(axes = var_4903_axes_0, x = linear_196_cast_fp16)[name = tensor<string, []>("op_4903_cast_fp16")];
            tensor<string, []> var_4903_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4903_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> var_4906_axis_0 = const()[name = tensor<string, []>("op_4906_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [28, 1, 4, 2048, 128]> var_4906_cast_fp16 = stack(axis = var_4906_axis_0, values = (k_full_1_cast_fp16, k_full_3_cast_fp16, k_full_5_cast_fp16, k_full_7_cast_fp16, k_full_9_cast_fp16, k_full_11_cast_fp16, k_full_13_cast_fp16, k_full_15_cast_fp16, k_full_17_cast_fp16, k_full_19_cast_fp16, k_full_21_cast_fp16, k_full_23_cast_fp16, k_full_25_cast_fp16, k_full_27_cast_fp16, k_full_29_cast_fp16, k_full_31_cast_fp16, k_full_33_cast_fp16, k_full_35_cast_fp16, k_full_37_cast_fp16, k_full_39_cast_fp16, k_full_41_cast_fp16, k_full_43_cast_fp16, k_full_45_cast_fp16, k_full_47_cast_fp16, k_full_49_cast_fp16, k_full_51_cast_fp16, k_full_53_cast_fp16, k_full_cast_fp16))[name = tensor<string, []>("op_4906_cast_fp16")];
            tensor<string, []> var_4906_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4906_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<int32, []> var_4909_axis_0 = const()[name = tensor<string, []>("op_4909_axis_0"), val = tensor<int32, []>(0)];
            tensor<fp16, [28, 1, 4, 2048, 128]> var_4909_cast_fp16 = stack(axis = var_4909_axis_0, values = (v_full_1_cast_fp16, v_full_3_cast_fp16, v_full_5_cast_fp16, v_full_7_cast_fp16, v_full_9_cast_fp16, v_full_11_cast_fp16, v_full_13_cast_fp16, v_full_15_cast_fp16, v_full_17_cast_fp16, v_full_19_cast_fp16, v_full_21_cast_fp16, v_full_23_cast_fp16, v_full_25_cast_fp16, v_full_27_cast_fp16, v_full_29_cast_fp16, v_full_31_cast_fp16, v_full_33_cast_fp16, v_full_35_cast_fp16, v_full_37_cast_fp16, v_full_39_cast_fp16, v_full_41_cast_fp16, v_full_43_cast_fp16, v_full_45_cast_fp16, v_full_47_cast_fp16, v_full_49_cast_fp16, v_full_51_cast_fp16, v_full_53_cast_fp16, v_full_cast_fp16))[name = tensor<string, []>("op_4909_cast_fp16")];
            tensor<string, []> var_4909_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("op_4909_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
            tensor<fp32, [28, 1, 4, 2048, 128]> kv_v_out = cast(dtype = var_4909_cast_fp16_to_fp32_dtype_0, x = var_4909_cast_fp16)[name = tensor<string, []>("cast_176")];
            tensor<fp32, [28, 1, 4, 2048, 128]> kv_k_out = cast(dtype = var_4906_cast_fp16_to_fp32_dtype_0, x = var_4906_cast_fp16)[name = tensor<string, []>("cast_177")];
            tensor<fp32, [1, 217232]> logits = cast(dtype = var_4903_cast_fp16_to_fp32_dtype_0, x = var_4903_cast_fp16)[name = tensor<string, []>("cast_178")];
        } -> (logits, kv_k_out, kv_v_out);
}