File size: 93,956 Bytes
7def60a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 |
// llama.cpp gRPC C++ backend server
//
// Ettore Di Giacinto <[email protected]> and llama.cpp authors
//
// This is a gRPC server for llama.cpp compatible with the LocalAI proto
// Note: this is a re-adaptation of the original llama.cpp example/server.cpp for HTTP (https://github.com/ggerganov/llama.cpp/tree/master/examples/server),
// but modified to work with gRPC
//
#include <iostream>
#include <memory>
#include <string>
#include <getopt.h>
#include "clip.h"
#include "llava.h"
#include "stb_image.h"
#include "common.h"
#include "json.hpp"
#include "llama.h"
#include "grammar-parser.h"
#include "backend.pb.h"
#include "backend.grpc.pb.h"
#include "utils.hpp"
// include std::regex
#include <cstddef>
#include <thread>
#include <mutex>
#include <chrono>
#include <regex>
#include <condition_variable>
#include <grpcpp/ext/proto_server_reflection_plugin.h>
#include <grpcpp/grpcpp.h>
#include <grpcpp/health_check_service_interface.h>
#include <atomic>
#include <signal.h>
using grpc::Server;
using grpc::ServerBuilder;
using grpc::ServerContext;
using grpc::Status;
using backend::HealthMessage;
///// LLAMA.CPP server code below
using json = nlohmann::json;
struct server_params
{
std::string hostname = "127.0.0.1";
std::vector<std::string> api_keys;
std::string public_path = "examples/server/public";
std::string chat_template = "";
int32_t port = 8080;
int32_t read_timeout = 600;
int32_t write_timeout = 600;
bool slots_endpoint = true;
bool metrics_endpoint = false;
};
bool server_verbose = false;
bool server_log_json = true;
static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
{
size_t i;
for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++)
{
}
return i;
}
enum stop_type
{
STOP_FULL,
STOP_PARTIAL,
};
static bool ends_with(const std::string &str, const std::string &suffix)
{
return str.size() >= suffix.size() &&
0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
}
static size_t find_partial_stop_string(const std::string &stop,
const std::string &text)
{
if (!text.empty() && !stop.empty())
{
const char text_last_char = text.back();
for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--)
{
if (stop[char_index] == text_last_char)
{
const std::string current_partial = stop.substr(0, char_index + 1);
if (ends_with(text, current_partial))
{
return text.size() - char_index - 1;
}
}
}
}
return std::string::npos;
}
// TODO: reuse llama_detokenize
template <class Iter>
static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end)
{
std::string ret;
for (; begin != end; ++begin)
{
ret += llama_token_to_piece(ctx, *begin);
}
return ret;
}
// format incomplete utf-8 multibyte character for output
static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token)
{
std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token);
// if the size is 1 and first bit is 1, meaning it's a partial character
// (size > 1 meaning it's already a known token)
if (out.size() == 1 && (out[0] & 0x80) == 0x80)
{
std::stringstream ss;
ss << std::hex << (out[0] & 0xff);
std::string res(ss.str());
out = "byte: \\x" + res;
}
return out;
}
// convert a vector of completion_token_output to json
static json probs_vector_to_json(const llama_context *ctx, const std::vector<completion_token_output> &probs)
{
json out = json::array();
for (const auto &prob : probs)
{
json probs_for_token = json::array();
for (const auto &p : prob.probs)
{
std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok);
probs_for_token.push_back(json
{
{"tok_str", tok_str},
{"prob", p.prob},
});
}
std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok);
out.push_back(json{
{"content", tok_str},
{"probs", probs_for_token},
});
}
return out;
}
struct llama_client_slot
{
int id;
int task_id = -1;
struct slot_params params;
slot_state state = IDLE;
slot_command command = NONE;
// used to determine the slot that has been used the longest
int64_t t_last_used = -1;
// generation props
int32_t n_ctx = 0; // context size per slot
int32_t n_past = 0;
int32_t n_decoded = 0;
int32_t n_remaining = -1;
int32_t i_batch = -1;
int32_t n_predict = -1;
int32_t num_prompt_tokens = 0;
int32_t num_prompt_tokens_processed = 0;
json prompt;
std::string generated_text;
llama_token sampled;
std::vector<llama_token> cache_tokens;
std::vector<completion_token_output> generated_token_probs;
bool infill = false;
bool embedding = false;
bool has_next_token = true;
bool truncated = false;
bool stopped_eos = false;
bool stopped_word = false;
bool stopped_limit = false;
bool oaicompat = false;
std::string oaicompat_model;
std::string stopping_word;
// sampling
struct llama_sampling_params sparams;
llama_sampling_context *ctx_sampling = nullptr;
int32_t ga_i = 0; // group-attention state
int32_t ga_n = 1; // group-attention factor
int32_t ga_w = 512; // group-attention width
int32_t n_past_se = 0; // self-extend
// multimodal
std::vector<slot_image> images;
// stats
size_t sent_count = 0;
size_t sent_token_probs_index = 0;
int64_t t_start_process_prompt;
int64_t t_start_genereration;
double t_prompt_processing; // ms
double t_token_generation; // ms
// multitasks
int multitask_id = -1;
void reset() {
num_prompt_tokens = 0;
generated_text = "";
truncated = false;
stopped_eos = false;
stopped_word = false;
stopped_limit = false;
stopping_word = "";
n_past = 0;
sent_count = 0;
sent_token_probs_index = 0;
infill = false;
ga_i = 0;
n_past_se = 0;
generated_token_probs.clear();
for (slot_image & img : images)
{
free(img.image_embedding);
if (img.img_data) {
clip_image_u8_free(img.img_data);
}
img.prefix_prompt = "";
}
images.clear();
}
bool has_budget(gpt_params &global_params) {
if (params.n_predict == -1 && global_params.n_predict == -1)
{
return true; // limitless
}
n_remaining = -1;
if (params.n_predict != -1)
{
n_remaining = params.n_predict - n_decoded;
}
else if (global_params.n_predict != -1)
{
n_remaining = global_params.n_predict - n_decoded;
}
return n_remaining > 0; // no budget
}
bool available() const {
return state == IDLE && command == NONE;
}
bool is_processing() const {
return (state == IDLE && command == LOAD_PROMPT) || state == PROCESSING;
}
void add_token_string(const completion_token_output &token) {
if (command == RELEASE)
{
return;
}
cache_tokens.push_back(token.tok);
generated_token_probs.push_back(token);
}
void release() {
if (state == PROCESSING)
{
t_token_generation = (ggml_time_us() - t_start_genereration) / 1e3;
command = RELEASE;
}
}
json get_formated_timings() {
return json
{
{"prompt_n", num_prompt_tokens_processed},
{"prompt_ms", t_prompt_processing},
{"prompt_per_token_ms", t_prompt_processing / num_prompt_tokens_processed},
{"prompt_per_second", 1e3 / t_prompt_processing * num_prompt_tokens_processed},
{"predicted_n", n_decoded},
{"predicted_ms", t_token_generation},
{"predicted_per_token_ms", t_token_generation / n_decoded},
{"predicted_per_second", 1e3 / t_token_generation * n_decoded},
};
}
void print_timings() const {
char buffer[512];
double t_token = t_prompt_processing / num_prompt_tokens_processed;
double n_tokens_second = 1e3 / t_prompt_processing * num_prompt_tokens_processed;
sprintf(buffer, "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
t_prompt_processing, num_prompt_tokens_processed,
t_token, n_tokens_second);
LOG_INFO(buffer, {
{"slot_id", id},
{"task_id", task_id},
{"t_prompt_processing", t_prompt_processing},
{"num_prompt_tokens_processed", num_prompt_tokens_processed},
{"t_token", t_token},
{"n_tokens_second", n_tokens_second},
});
t_token = t_token_generation / n_decoded;
n_tokens_second = 1e3 / t_token_generation * n_decoded;
sprintf(buffer, "generation eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)",
t_token_generation, n_decoded,
t_token, n_tokens_second);
LOG_INFO(buffer, {
{"slot_id", id},
{"task_id", task_id},
{"t_token_generation", t_token_generation},
{"n_decoded", n_decoded},
{"t_token", t_token},
{"n_tokens_second", n_tokens_second},
});
sprintf(buffer, " total time = %10.2f ms", t_prompt_processing + t_token_generation);
LOG_INFO(buffer, {
{"slot_id", id},
{"task_id", task_id},
{"t_prompt_processing", t_prompt_processing},
{"t_token_generation", t_token_generation},
{"t_total", t_prompt_processing + t_token_generation},
});
}
};
struct llama_metrics {
uint64_t n_prompt_tokens_processed_total = 0;
uint64_t n_tokens_predicted_total = 0;
uint64_t n_prompt_tokens_processed = 0;
uint64_t t_prompt_processing = 0;
uint64_t n_tokens_predicted = 0;
uint64_t t_tokens_generation = 0;
void on_prompt_eval(const llama_client_slot &slot) {
n_prompt_tokens_processed_total += slot.num_prompt_tokens_processed;
n_prompt_tokens_processed += slot.num_prompt_tokens_processed;
t_prompt_processing += slot.t_prompt_processing;
}
void on_prediction(const llama_client_slot &slot) {
n_tokens_predicted_total += slot.n_decoded;
n_tokens_predicted += slot.n_decoded;
t_tokens_generation += slot.t_token_generation;
}
void reset_bucket() {
n_prompt_tokens_processed = 0;
t_prompt_processing = 0;
n_tokens_predicted = 0;
t_tokens_generation = 0;
}
};
struct llama_server_context
{
llama_model *model = nullptr;
llama_context *ctx = nullptr;
clip_ctx *clp_ctx = nullptr;
gpt_params params;
llama_batch batch;
bool multimodal = false;
bool clean_kv_cache = true;
bool all_slots_are_idle = false;
bool add_bos_token = true;
int32_t n_ctx; // total context for all clients / slots
// system prompt
bool system_need_update = false;
std::string system_prompt;
std::vector<llama_token> system_tokens;
std::string name_user; // this should be the antiprompt
std::string name_assistant;
// slots / clients
std::vector<llama_client_slot> slots;
json default_generation_settings_for_props;
llama_server_queue queue_tasks;
llama_server_response queue_results;
llama_metrics metrics;
~llama_server_context()
{
if (ctx)
{
llama_free(ctx);
ctx = nullptr;
}
if (model)
{
llama_free_model(model);
model = nullptr;
}
}
bool load_model(const gpt_params ¶ms_)
{
params = params_;
if (!params.mmproj.empty()) {
multimodal = true;
LOG_INFO("Multi Modal Mode Enabled", {});
clp_ctx = clip_model_load(params.mmproj.c_str(), /*verbosity=*/ 1);
if(clp_ctx == nullptr) {
LOG_ERROR("unable to load clip model", {{"model", params.mmproj}});
return false;
}
if (params.n_ctx < 2048) { // request larger context for the image embedding
params.n_ctx = 2048;
}
}
llama_init_result llama_init = llama_init_from_gpt_params(params);
model = llama_init.model;
ctx = llama_init.context;
if (model == nullptr)
{
LOG_ERROR("unable to load model", {{"model", params.model}});
return false;
}
if (multimodal) {
const int n_embd_clip = clip_n_mmproj_embd(clp_ctx);
const int n_embd_llm = llama_n_embd(model);
if (n_embd_clip != n_embd_llm) {
LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_embd_clip, n_embd_llm);
llama_free(ctx);
llama_free_model(model);
return false;
}
}
n_ctx = llama_n_ctx(ctx);
add_bos_token = llama_should_add_bos_token(model);
return true;
}
void validate_model_chat_template(server_params & sparams) {
llama_chat_message chat[] = {{"user", "test"}};
std::vector<char> buf(1);
int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size());
if (res < 0) {
LOG_ERROR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {});
sparams.chat_template = "<|im_start|>"; // llama_chat_apply_template only checks if <|im_start|> exist in the template
}
}
void initialize() {
// create slots
all_slots_are_idle = true;
const int32_t n_ctx_slot = n_ctx / params.n_parallel;
LOG_INFO("initializing slots", {{"n_slots", params.n_parallel}});
for (int i = 0; i < params.n_parallel; i++)
{
llama_client_slot slot;
slot.id = i;
slot.n_ctx = n_ctx_slot;
slot.n_predict = params.n_predict;
LOG_INFO("new slot", {
{"slot_id", slot.id},
{"n_ctx_slot", slot.n_ctx}
});
const int ga_n = params.grp_attn_n;
const int ga_w = params.grp_attn_w;
if (ga_n != 1) {
GGML_ASSERT(ga_n > 0 && "ga_n must be positive"); // NOLINT
GGML_ASSERT(ga_w % ga_n == 0 && "ga_w must be a multiple of ga_n"); // NOLINT
//GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of ga_w"); // NOLINT
//GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * ga_n"); // NOLINT
LOG_INFO("slot self-extend", {
{"slot_id", slot.id},
{"ga_n", ga_n},
{"ga_w", ga_w}
});
}
slot.ga_i = 0;
slot.ga_n = ga_n;
slot.ga_w = ga_w;
slot.reset();
slots.push_back(slot);
}
default_generation_settings_for_props = get_formated_generation(slots.front());
default_generation_settings_for_props["seed"] = -1;
batch = llama_batch_init(n_ctx, 0, params.n_parallel);
}
std::vector<llama_token> tokenize(const json & json_prompt, bool add_bos) const
{
// TODO: currently, we tokenize using special tokens by default
// this is not always correct (see https://github.com/ggerganov/llama.cpp/pull/4160#issuecomment-1824826216)
// but it's better compared to completely ignoring ChatML and other chat templates
const bool TMP_FORCE_SPECIAL = true;
// If `add_bos` is true, we only add BOS, when json_prompt is a string,
// or the first element of the json_prompt array is a string.
std::vector<llama_token> prompt_tokens;
if (json_prompt.is_array())
{
bool first = true;
for (const auto& p : json_prompt)
{
if (p.is_string())
{
auto s = p.template get<std::string>();
std::vector<llama_token> p;
if (first)
{
p = ::llama_tokenize(ctx, s, add_bos, TMP_FORCE_SPECIAL);
first = false;
}
else
{
p = ::llama_tokenize(ctx, s, false, TMP_FORCE_SPECIAL);
}
prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
}
else
{
if (first)
{
first = false;
}
prompt_tokens.push_back(p.template get<llama_token>());
}
}
}
else
{
auto s = json_prompt.template get<std::string>();
prompt_tokens = ::llama_tokenize(ctx, s, add_bos, TMP_FORCE_SPECIAL);
}
return prompt_tokens;
}
llama_client_slot* get_slot(int id) {
int64_t t_last = ggml_time_us();
llama_client_slot *last_used = nullptr;
for (llama_client_slot & slot : slots)
{
if (slot.id == id && slot.available())
{
return &slot;
}
if (slot.available() && slot.t_last_used < t_last)
{
last_used = &slot;
t_last = slot.t_last_used;
}
}
return last_used;
}
bool launch_slot_with_data(llama_client_slot* &slot, json data) {
slot_params default_params;
llama_sampling_params default_sparams;
slot->params.stream = json_value(data, "stream", false);
slot->params.cache_prompt = json_value(data, "cache_prompt", false);
slot->params.n_predict = json_value(data, "n_predict", default_params.n_predict);
slot->sparams.top_k = json_value(data, "top_k", default_sparams.top_k);
slot->sparams.top_p = json_value(data, "top_p", default_sparams.top_p);
slot->sparams.min_p = json_value(data, "min_p", default_sparams.min_p);
slot->sparams.tfs_z = json_value(data, "tfs_z", default_sparams.tfs_z);
slot->sparams.typical_p = json_value(data, "typical_p", default_sparams.typical_p);
slot->sparams.temp = json_value(data, "temperature", default_sparams.temp);
slot->sparams.dynatemp_range = json_value(data, "dynatemp_range", default_sparams.dynatemp_range);
slot->sparams.dynatemp_exponent = json_value(data, "dynatemp_exponent", default_sparams.dynatemp_exponent);
slot->sparams.penalty_last_n = json_value(data, "repeat_last_n", default_sparams.penalty_last_n);
slot->sparams.penalty_repeat = json_value(data, "repeat_penalty", default_sparams.penalty_repeat);
slot->sparams.penalty_freq = json_value(data, "frequency_penalty", default_sparams.penalty_freq);
slot->sparams.penalty_present = json_value(data, "presence_penalty", default_sparams.penalty_present);
slot->sparams.mirostat = json_value(data, "mirostat", default_sparams.mirostat);
slot->sparams.mirostat_tau = json_value(data, "mirostat_tau", default_sparams.mirostat_tau);
slot->sparams.mirostat_eta = json_value(data, "mirostat_eta", default_sparams.mirostat_eta);
slot->sparams.penalize_nl = json_value(data, "penalize_nl", default_sparams.penalize_nl);
slot->params.n_keep = json_value(data, "n_keep", slot->params.n_keep);
slot->params.seed = json_value(data, "seed", default_params.seed);
slot->sparams.grammar = json_value(data, "grammar", default_sparams.grammar);
slot->sparams.n_probs = json_value(data, "n_probs", default_sparams.n_probs);
slot->sparams.min_keep = json_value(data, "min_keep", default_sparams.min_keep);
if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) {
// Might be better to reject the request with a 400 ?
LOG_WARNING("Max tokens to predict exceeds server configuration", {
{"params.n_predict", slot->params.n_predict},
{"slot.n_predict", slot->n_predict},
});
slot->params.n_predict = slot->n_predict;
}
// infill
if (data.count("input_prefix") != 0)
{
slot->params.input_prefix = data["input_prefix"];
}
else
{
slot->params.input_prefix = "";
}
if (data.count("input_suffix") != 0)
{
slot->params.input_suffix = data["input_suffix"];
}
else
{
slot->params.input_suffix = "";
}
if (data.count("prompt") != 0)
{
slot->prompt = data["prompt"];
}
else
{
slot->prompt = "";
}
slot->sparams.penalty_prompt_tokens.clear();
slot->sparams.use_penalty_prompt_tokens = false;
const auto &penalty_prompt = data.find("penalty_prompt");
if (penalty_prompt != data.end())
{
if (penalty_prompt->is_string())
{
const auto penalty_prompt_string = penalty_prompt->get<std::string>();
auto penalty_tokens = llama_tokenize(model, penalty_prompt_string, false);
slot->sparams.penalty_prompt_tokens.swap(penalty_tokens);
if (slot->params.n_predict > 0)
{
slot->sparams.penalty_prompt_tokens.reserve(slot->sparams.penalty_prompt_tokens.size() + slot->params.n_predict);
}
slot->sparams.use_penalty_prompt_tokens = true;
}
else if (penalty_prompt->is_array())
{
const auto n_tokens = penalty_prompt->size();
slot->sparams.penalty_prompt_tokens.reserve(n_tokens + std::max(0, slot->params.n_predict));
const int n_vocab = llama_n_vocab(model);
for (const auto &penalty_token : *penalty_prompt)
{
if (penalty_token.is_number_integer())
{
const auto tok = penalty_token.get<llama_token>();
if (tok >= 0 && tok < n_vocab)
{
slot->sparams.penalty_prompt_tokens.push_back(tok);
}
}
}
slot->sparams.use_penalty_prompt_tokens = true;
}
}
slot->sparams.logit_bias.clear();
if (json_value(data, "ignore_eos", false))
{
slot->sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
}
const auto &logit_bias = data.find("logit_bias");
if (logit_bias != data.end() && logit_bias->is_array())
{
const int n_vocab = llama_n_vocab(model);
for (const auto &el : *logit_bias)
{
if (el.is_array() && el.size() == 2)
{
float bias;
if (el[1].is_number())
{
bias = el[1].get<float>();
}
else if (el[1].is_boolean() && !el[1].get<bool>())
{
bias = -INFINITY;
}
else
{
continue;
}
if (el[0].is_number_integer())
{
llama_token tok = el[0].get<llama_token>();
if (tok >= 0 && tok < n_vocab)
{
slot->sparams.logit_bias[tok] = bias;
}
}
else if (el[0].is_string())
{
auto toks = llama_tokenize(model, el[0].get<std::string>(), false);
for (auto tok : toks)
{
slot->sparams.logit_bias[tok] = bias;
}
}
}
}
}
slot->params.antiprompt.clear();
const auto &stop = data.find("stop");
if (stop != data.end() && stop->is_array())
{
for (const auto &word : *stop)
{
if (!word.empty())
{
slot->params.antiprompt.push_back(word);
}
}
}
const auto &samplers_sequence = data.find("samplers");
if (samplers_sequence != data.end() && samplers_sequence->is_array())
{
std::vector<std::string> sampler_names;
for (const auto &sampler_name : *samplers_sequence)
{
if (sampler_name.is_string())
{
sampler_names.emplace_back(sampler_name);
}
}
slot->sparams.samplers_sequence = llama_sampling_types_from_names(sampler_names, false);
}
else
{
slot->sparams.samplers_sequence = default_sparams.samplers_sequence;
}
if (multimodal)
{
const auto &images_data = data.find("image_data");
if (images_data != data.end() && images_data->is_array())
{
for (const auto &img : *images_data)
{
const std::vector<uint8_t> image_buffer = base64_decode(img["data"].get<std::string>());
slot_image img_sl;
img_sl.id = img.count("id") != 0 ? img["id"].get<int>() : slot->images.size();
img_sl.img_data = clip_image_u8_init();
if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data))
{
LOG_ERROR("failed to load image", {
{"slot_id", slot->id},
{"img_sl_id", img_sl.id}
});
return false;
}
LOG_VERBOSE("image loaded", {
{"slot_id", slot->id},
{"img_sl_id", img_sl.id}
});
img_sl.request_encode_image = true;
slot->images.push_back(img_sl);
}
// process prompt
// example: system prompt [img-102] user [img-103] describe [img-134] -> [{id: 102, prefix: 'system prompt '}, {id: 103, prefix: ' user '}, {id: 134, prefix: ' describe '}]}
if (slot->images.size() > 0 && !slot->prompt.is_array())
{
std::string prompt = slot->prompt.get<std::string>();
size_t pos = 0, begin_prefix = 0;
std::string pattern = "[img-";
while ((pos = prompt.find(pattern, pos)) != std::string::npos) {
size_t end_prefix = pos;
pos += pattern.length();
size_t end_pos = prompt.find(']', pos);
if (end_pos != std::string::npos)
{
std::string image_id = prompt.substr(pos, end_pos - pos);
try
{
int img_id = std::stoi(image_id);
bool found = false;
for (slot_image &img : slot->images)
{
if (img.id == img_id) {
found = true;
img.prefix_prompt = prompt.substr(begin_prefix, end_prefix - begin_prefix);
begin_prefix = end_pos + 1;
break;
}
}
if (!found) {
LOG_TEE("ERROR: Image with id: %i, not found.\n", img_id);
slot->images.clear();
return false;
}
} catch (const std::invalid_argument& e) {
LOG_TEE("Invalid image number id in prompt\n");
slot->images.clear();
return false;
}
}
}
slot->prompt = "";
slot->params.input_suffix = prompt.substr(begin_prefix);
slot->params.cache_prompt = false; // multimodal doesn't support cache prompt
}
}
}
if (slot->ctx_sampling != nullptr)
{
llama_sampling_free(slot->ctx_sampling);
}
slot->ctx_sampling = llama_sampling_init(slot->sparams);
llama_set_rng_seed(ctx, slot->params.seed);
slot->command = LOAD_PROMPT;
all_slots_are_idle = false;
LOG_INFO("slot is processing task", {
{"slot_id", slot->id},
{"task_id", slot->task_id},
});
LOG_TEE("sampling: \n%s\n", llama_sampling_print(slot->sparams).c_str());
return true;
}
void kv_cache_clear() {
// clear the entire KV cache
llama_kv_cache_clear(ctx);
clean_kv_cache = false;
}
void update_system_prompt() {
kv_cache_clear();
system_tokens.clear();
if (!system_prompt.empty()) {
system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token);
llama_batch_clear(batch);
for (int i = 0; i < (int)system_tokens.size(); ++i)
{
llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
}
for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += params.n_batch)
{
const int32_t n_tokens = std::min(params.n_batch, (int32_t) (batch.n_tokens - i));
llama_batch batch_view = {
n_tokens,
batch.token + i,
nullptr,
batch.pos + i,
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
if (llama_decode(ctx, batch_view) != 0)
{
LOG_TEE("%s: llama_decode() failed\n", __func__);
return;
}
}
// assign the system KV cache to all parallel sequences
for (int32_t i = 1; i < params.n_parallel; ++i)
{
llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size());
}
}
LOG_TEE("system prompt updated\n");
system_need_update = false;
}
void notify_system_prompt_changed() {
// release all slots
for (llama_client_slot &slot : slots)
{
slot.release();
}
system_need_update = true;
}
void process_system_prompt_data(const json &sys_props) {
system_prompt = sys_props.value("prompt", "");
name_user = sys_props.value("anti_prompt", "");
name_assistant = sys_props.value("assistant_name", "");
notify_system_prompt_changed();
}
static size_t find_stopping_strings(const std::string &text, const size_t last_token_size,
const stop_type type, llama_client_slot &slot)
{
size_t stop_pos = std::string::npos;
for (const std::string &word : slot.params.antiprompt)
{
size_t pos;
if (type == STOP_FULL)
{
const size_t tmp = word.size() + last_token_size;
const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0;
pos = text.find(word, from_pos);
}
else
{
pos = find_partial_stop_string(word, text);
}
if (pos != std::string::npos &&
(stop_pos == std::string::npos || pos < stop_pos))
{
if (type == STOP_FULL)
{
slot.stopped_word = true;
slot.stopping_word = word;
slot.has_next_token = false;
}
stop_pos = pos;
}
}
return stop_pos;
}
bool process_token(completion_token_output &result, llama_client_slot &slot) {
// remember which tokens were sampled - used for repetition penalties during sampling
const std::string token_str = llama_token_to_piece(ctx, result.tok);
slot.sampled = result.tok;
// search stop word and delete it
slot.generated_text += token_str;
slot.has_next_token = true;
if (slot.ctx_sampling->params.use_penalty_prompt_tokens && result.tok != -1)
{
// we can change penalty_prompt_tokens because it is always created from scratch each request
slot.ctx_sampling->params.penalty_prompt_tokens.push_back(result.tok);
}
// check if there is incomplete UTF-8 character at the end
bool incomplete = false;
for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i)
{
unsigned char c = slot.generated_text[slot.generated_text.size() - i];
if ((c & 0xC0) == 0x80)
{
// continuation byte: 10xxxxxx
continue;
}
if ((c & 0xE0) == 0xC0)
{
// 2-byte character: 110xxxxx ...
incomplete = i < 2;
}
else if ((c & 0xF0) == 0xE0)
{
// 3-byte character: 1110xxxx ...
incomplete = i < 3;
}
else if ((c & 0xF8) == 0xF0)
{
// 4-byte character: 11110xxx ...
incomplete = i < 4;
}
// else 1-byte character or invalid byte
break;
}
if (!incomplete)
{
size_t pos = std::min(slot.sent_count, slot.generated_text.size());
const std::string str_test = slot.generated_text.substr(pos);
bool is_stop_full = false;
size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot);
if (stop_pos != std::string::npos)
{
is_stop_full = true;
slot.generated_text.erase(
slot.generated_text.begin() + pos + stop_pos,
slot.generated_text.end());
pos = std::min(slot.sent_count, slot.generated_text.size());
}
else
{
is_stop_full = false;
stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_PARTIAL, slot);
}
// check if there is any token to predict
if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0))
{
// no send the stop word in the response
result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
slot.sent_count += result.text_to_send.size();
// add the token to slot queue and cache
}
slot.add_token_string(result);
if (slot.params.stream)
{
send_partial_response(slot, result);
}
}
if (incomplete)
{
slot.has_next_token = true;
}
// check the limits
if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params))
{
slot.stopped_limit = true;
slot.has_next_token = false;
}
if (result.tok == llama_token_eos(model))
{
slot.stopped_eos = true;
slot.has_next_token = false;
LOG_VERBOSE("eos token found", {});
}
LOG_VERBOSE("next token", {
{"token", result.tok},
{"token_text", tokens_to_output_formatted_string(ctx, result.tok)},
{"has_next_token", slot.has_next_token},
{"n_remain", slot.n_remaining},
{"num_tokens_predicted", slot.n_decoded},
{"stopped_eos", slot.stopped_eos},
{"stopped_word", slot.stopped_word},
{"stopped_limit", slot.stopped_limit},
{"stopping_word", slot.stopping_word},
});
return slot.has_next_token; // continue
}
bool process_images(llama_client_slot &slot) const
{
for (slot_image &img : slot.images)
{
if (!img.request_encode_image)
{
continue;
}
if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
LOG_TEE("Error processing the given image");
return false;
}
img.request_encode_image = false;
}
return slot.images.size() > 0;
}
void send_error(task_server& task, const std::string &error)
{
LOG_TEE("task %i - error: %s\n", task.id, error.c_str());
task_result res;
res.id = task.id;
res.multitask_id = task.multitask_id;
res.stop = false;
res.error = true;
res.result_json = { { "content", error } };
queue_results.send(res);
}
json get_formated_generation(llama_client_slot &slot)
{
const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(model));
const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() &&
eos_bias->second < 0.0f && std::isinf(eos_bias->second);
std::vector<std::string> samplers_sequence;
for (const auto &sampler_type : slot.sparams.samplers_sequence)
{
samplers_sequence.emplace_back(llama_sampling_type_to_str(sampler_type));
}
return json {
{"n_ctx", slot.n_ctx},
{"n_predict", slot.n_predict},
{"model", params.model_alias},
{"seed", slot.params.seed},
{"temperature", slot.sparams.temp},
{"dynatemp_range", slot.sparams.dynatemp_range},
{"dynatemp_exponent", slot.sparams.dynatemp_exponent},
{"top_k", slot.sparams.top_k},
{"top_p", slot.sparams.top_p},
{"min_p", slot.sparams.min_p},
{"tfs_z", slot.sparams.tfs_z},
{"typical_p", slot.sparams.typical_p},
{"repeat_last_n", slot.sparams.penalty_last_n},
{"repeat_penalty", slot.sparams.penalty_repeat},
{"presence_penalty", slot.sparams.penalty_present},
{"frequency_penalty", slot.sparams.penalty_freq},
{"penalty_prompt_tokens", slot.sparams.penalty_prompt_tokens},
{"use_penalty_prompt_tokens", slot.sparams.use_penalty_prompt_tokens},
{"mirostat", slot.sparams.mirostat},
{"mirostat_tau", slot.sparams.mirostat_tau},
{"mirostat_eta", slot.sparams.mirostat_eta},
{"penalize_nl", slot.sparams.penalize_nl},
{"stop", slot.params.antiprompt},
{"n_predict", slot.params.n_predict},
{"n_keep", params.n_keep},
{"ignore_eos", ignore_eos},
{"stream", slot.params.stream},
{"logit_bias", slot.sparams.logit_bias},
{"n_probs", slot.sparams.n_probs},
{"min_keep", slot.sparams.min_keep},
{"grammar", slot.sparams.grammar},
{"samplers", samplers_sequence}
};
}
void send_partial_response(llama_client_slot &slot, completion_token_output tkn)
{
task_result res;
res.id = slot.task_id;
res.multitask_id = slot.multitask_id;
res.error = false;
res.stop = false;
res.result_json = json
{
{"content", tkn.text_to_send},
{"stop", false},
{"slot_id", slot.id},
{"multimodal", multimodal}
};
if (slot.sparams.n_probs > 0)
{
std::vector<completion_token_output> probs_output = {};
const std::vector<llama_token> to_send_toks = llama_tokenize(ctx, tkn.text_to_send, false);
size_t probs_pos = std::min(slot.sent_token_probs_index, slot.generated_token_probs.size());
size_t probs_stop_pos = std::min(slot.sent_token_probs_index + to_send_toks.size(), slot.generated_token_probs.size());
if (probs_pos < probs_stop_pos)
{
probs_output = std::vector<completion_token_output>(slot.generated_token_probs.begin() + probs_pos, slot.generated_token_probs.begin() + probs_stop_pos);
}
slot.sent_token_probs_index = probs_stop_pos;
res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs_output);
}
if (slot.oaicompat)
{
res.result_json["oaicompat_token_ctr"] = slot.n_decoded;
res.result_json["model"] = slot.oaicompat_model;
}
queue_results.send(res);
}
void send_final_response(llama_client_slot &slot)
{
task_result res;
res.id = slot.task_id;
res.multitask_id = slot.multitask_id;
res.error = false;
res.stop = true;
res.result_json = json
{
{"content", !slot.params.stream ? slot.generated_text : ""},
{"slot_id", slot.id},
{"stop", true},
{"model", params.model_alias},
{"tokens_predicted", slot.n_decoded},
{"tokens_evaluated", slot.num_prompt_tokens},
{"generation_settings", get_formated_generation(slot)},
{"prompt", slot.prompt},
{"truncated", slot.truncated},
{"stopped_eos", slot.stopped_eos},
{"stopped_word", slot.stopped_word},
{"stopped_limit", slot.stopped_limit},
{"stopping_word", slot.stopping_word},
{"tokens_cached", slot.n_past},
{"timings", slot.get_formated_timings()}
};
if (slot.sparams.n_probs > 0)
{
std::vector<completion_token_output> probs = {};
if (!slot.params.stream && slot.stopped_word)
{
const std::vector<llama_token> stop_word_toks = llama_tokenize(ctx, slot.stopping_word, false);
probs = std::vector<completion_token_output>(slot.generated_token_probs.begin(), slot.generated_token_probs.end() - stop_word_toks.size());
}
else
{
probs = std::vector<completion_token_output>(
slot.generated_token_probs.begin(),
slot.generated_token_probs.end());
}
res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs);
}
if (slot.oaicompat)
{
res.result_json["oaicompat_token_ctr"] = slot.n_decoded;
res.result_json["model"] = slot.oaicompat_model;
}
queue_results.send(res);
}
void send_embedding(llama_client_slot &slot)
{
task_result res;
res.id = slot.task_id;
res.multitask_id = slot.multitask_id;
res.error = false;
res.stop = true;
const int n_embd = llama_n_embd(model);
if (!params.embedding)
{
LOG_WARNING("embedding disabled", {
{"params.embedding", params.embedding},
});
res.result_json = json
{
{"embedding", std::vector<float>(n_embd, 0.0f)},
};
}
else
{
const float *data = llama_get_embeddings(ctx);
std::vector<float> embedding(data, data + n_embd);
res.result_json = json
{
{"embedding", embedding },
};
}
queue_results.send(res);
}
void request_completion(int task_id, json data, bool infill, bool embedding, int multitask_id)
{
task_server task;
task.id = task_id;
task.target_id = 0;
task.data = std::move(data);
task.infill_mode = infill;
task.embedding_mode = embedding;
task.type = TASK_TYPE_COMPLETION;
task.multitask_id = multitask_id;
// when a completion task's prompt array is not a singleton, we split it into multiple requests
// otherwise, it's a single-prompt task, we actually queue it
// if there's numbers in the prompt array it will be treated as an array of tokens
if (task.data.count("prompt") != 0 && task.data.at("prompt").size() > 1) {
bool numbers = false;
for (const auto& e : task.data.at("prompt")) {
if (e.is_number()) {
numbers = true;
break;
}
}
// NOTE: split_multiprompt_task() does not handle a mix of strings and numbers,
// it will completely stall the server. I don't know where the bug for this is.
//
// if there are numbers, it needs to be treated like a single prompt,
// queue_tasks handles a mix of strings and numbers just fine.
if (numbers) {
queue_tasks.post(task);
} else {
split_multiprompt_task(task_id, task);
}
} else {
queue_tasks.post(task);
}
}
// for multiple images processing
bool ingest_images(llama_client_slot &slot, int n_batch)
{
int image_idx = 0;
while (image_idx < (int) slot.images.size())
{
slot_image &img = slot.images[image_idx];
// process prefix prompt
for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch)
{
const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
llama_batch batch_view = {
n_tokens,
batch.token + i,
nullptr,
batch.pos + i,
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
if (llama_decode(ctx, batch_view))
{
LOG_TEE("%s : failed to eval\n", __func__);
return false;
}
}
// process image with llm
for (int i = 0; i < img.image_tokens; i += n_batch)
{
int n_eval = img.image_tokens - i;
if (n_eval > n_batch)
{
n_eval = n_batch;
}
const int n_embd = llama_n_embd(model);
llama_batch batch_img = { n_eval, nullptr, (img.image_embedding + i * n_embd), nullptr, nullptr, nullptr, nullptr, slot.n_past, 1, 0, };
if (llama_decode(ctx, batch_img))
{
LOG_TEE("%s : failed to eval image\n", __func__);
return false;
}
slot.n_past += n_eval;
}
image_idx++;
llama_batch_clear(batch);
// append prefix of next image
const auto json_prompt = (image_idx >= (int) slot.images.size()) ?
slot.params.input_suffix : // no more images, then process suffix prompt
(json)(slot.images[image_idx].prefix_prompt);
std::vector<llama_token> append_tokens = tokenize(json_prompt, false); // has next image
for (int i = 0; i < (int) append_tokens.size(); ++i)
{
llama_batch_add(batch, append_tokens[i], system_tokens.size() + slot.n_past, { slot.id }, true);
slot.n_past += 1;
}
}
return true;
}
void request_cancel(int task_id)
{
task_server task;
task.type = TASK_TYPE_CANCEL;
task.target_id = task_id;
queue_tasks.post(task);
}
void split_multiprompt_task(int multitask_id, task_server& multiprompt_task)
{
int prompt_count = multiprompt_task.data.at("prompt").size();
if (prompt_count <= 1) {
send_error(multiprompt_task, "error while handling multiple prompts");
return;
}
// generate all the ID for subtask
std::vector<int> subtask_ids(prompt_count);
for (int i = 0; i < prompt_count; i++)
{
subtask_ids[i] = queue_tasks.get_new_id();
}
// queue up the multitask so we can track its subtask progression
queue_tasks.add_multitask(multitask_id, subtask_ids);
// add subtasks
for (int i = 0; i < prompt_count; i++)
{
json subtask_data = multiprompt_task.data;
subtask_data["prompt"] = subtask_data["prompt"][i];
// subtasks inherit everything else (infill mode, embedding mode, etc.)
request_completion(subtask_ids[i], subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id);
}
}
void process_single_task(task_server& task)
{
switch (task.type)
{
case TASK_TYPE_COMPLETION: {
llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
if (slot == nullptr)
{
// if no slot is available, we defer this task for processing later
LOG_VERBOSE("no slot is available", {{"task_id", task.id}});
queue_tasks.defer(task);
break;
}
if (task.data.contains("system_prompt"))
{
if (!all_slots_are_idle) {
send_error(task, "system prompt can only be updated when all slots are idle");
break;
}
process_system_prompt_data(task.data["system_prompt"]);
// reset cache_tokens for all slots
for (llama_client_slot &slot : slots)
{
slot.cache_tokens.clear();
slot.n_past = 0;
slot.n_past_se = 0;
}
}
slot->reset();
slot->infill = task.infill_mode;
slot->embedding = task.embedding_mode;
slot->task_id = task.id;
slot->multitask_id = task.multitask_id;
if (!launch_slot_with_data(slot, task.data))
{
// send error result
send_error(task, "internal_error");
break;
}
} break;
case TASK_TYPE_CANCEL: { // release slot linked with the task id
for (auto & slot : slots)
{
if (slot.task_id == task.target_id)
{
slot.release();
break;
}
}
} break;
case TASK_TYPE_NEXT_RESPONSE: {
// do nothing
} break;
}
}
void on_finish_multitask(task_multi& multitask)
{
// all subtasks done == multitask is done
task_result result;
result.id = multitask.id;
result.stop = true;
result.error = false;
// collect json results into one json result
std::vector<json> result_jsons;
for (auto& subres : multitask.results)
{
result_jsons.push_back(subres.result_json);
result.error = result.error && subres.error;
}
result.result_json = json{ { "results", result_jsons } };
queue_results.send(result);
}
bool update_slots() {
if (system_need_update)
{
LOG_INFO("updating system prompt", {});
update_system_prompt();
}
llama_batch_clear(batch);
if (all_slots_are_idle)
{
if (system_prompt.empty() && clean_kv_cache)
{
LOG_INFO("all slots are idle and system prompt is empty, clear the KV cache", {});
kv_cache_clear();
}
return true;
}
LOG_VERBOSE("posting NEXT_RESPONSE", {});
task_server task;
task.type = TASK_TYPE_NEXT_RESPONSE;
task.target_id = -1;
queue_tasks.post(task);
for (llama_client_slot &slot : slots)
{
if (slot.ga_n == 1)
{
if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx)
{
// START LOCALAI changes
// Temporary disable context-shifting as it can lead to infinite loops (issue: https://github.com/ggerganov/llama.cpp/issues/3969)
// See: https://github.com/mudler/LocalAI/issues/1333
// Context is exhausted, release the slot
slot.release();
send_final_response(slot);
slot.cache_tokens.clear();
slot.n_past = 0;
slot.truncated = false;
slot.has_next_token = true;
LOG_TEE("Context exhausted. Slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size());
continue;
// END LOCALAI changes
}
}
}
// decode any currently ongoing sequences
LOG_VERBOSE("decoding ongoing sequences", {});
for (auto & slot : slots)
{
// release the slot
if (slot.command == RELEASE)
{
slot.state = IDLE;
slot.command = NONE;
slot.t_last_used = ggml_time_us();
LOG_INFO("slot released", {
{"slot_id", slot.id},
{"task_id", slot.task_id},
{"n_ctx", n_ctx},
{"n_past", slot.n_past},
{"n_system_tokens", system_tokens.size()},
{"n_cache_tokens", slot.cache_tokens.size()},
{"truncated", slot.truncated}
});
queue_tasks.notify_slot_changed();
continue;
}
if (slot.state == IDLE)
{
continue;
}
slot.i_batch = batch.n_tokens;
const int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past;
// TODO: we always have to take into account the "system_tokens"
// this is not great and needs to be improved somehow
llama_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id }, true);
slot.n_past += 1;
}
// process in chunks of params.n_batch
int32_t n_batch = params.n_batch;
// assign workload to the slots
if (params.cont_batching || batch.n_tokens == 0)
{
for (auto & slot : slots)
{
const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty()) || !slot.images.empty();
// empty prompt passed -> release the slot and send empty response
// note: infill mode allows empty prompt
if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt && !slot.infill)
{
slot.release();
slot.print_timings();
send_final_response(slot);
continue;
}
// need process the prompt
if (slot.state == IDLE && slot.command == LOAD_PROMPT)
{
slot.state = PROCESSING;
slot.command = NONE;
std::vector<llama_token> prompt_tokens;
slot.t_start_process_prompt = ggml_time_us();
slot.t_start_genereration = 0;
if (slot.infill)
{
bool suff_rm_leading_spc = true;
if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1)
{
params.input_suffix.erase(0, 1);
suff_rm_leading_spc = false;
}
auto prefix_tokens = tokenize(slot.params.input_prefix, false);
auto suffix_tokens = tokenize(slot.params.input_suffix, false);
const int space_token = 29871; // TODO: this should not be hardcoded
if (suff_rm_leading_spc && !suffix_tokens.empty() && suffix_tokens[0] == space_token) {
suffix_tokens.erase(suffix_tokens.begin());
}
prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model));
prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(model));
prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
prefix_tokens.push_back(llama_token_middle(model));
prompt_tokens = prefix_tokens;
}
else
{
prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token); // add BOS if there isn't system prompt
}
slot.num_prompt_tokens = prompt_tokens.size();
if (slot.params.n_keep < 0)
{
slot.params.n_keep = slot.num_prompt_tokens;
}
slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep);
// if input prompt is too big, truncate it
if (slot.num_prompt_tokens >= slot.n_ctx)
{
const int n_left = slot.n_ctx - slot.params.n_keep;
const int n_block_size = n_left / 2;
const int erased_blocks = (slot.num_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size;
std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + slot.params.n_keep);
new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, prompt_tokens.end());
LOG_VERBOSE("input truncated", {
{"n_ctx", slot.n_ctx},
{"n_keep", slot.params.n_keep},
{"n_left", n_left},
{"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
});
slot.truncated = true;
prompt_tokens = new_tokens;
slot.num_prompt_tokens = prompt_tokens.size();
GGML_ASSERT(slot.num_prompt_tokens < slot.n_ctx);
}
if (!slot.params.cache_prompt)
{
llama_sampling_reset(slot.ctx_sampling);
slot.n_past = 0;
slot.n_past_se = 0;
slot.ga_i = 0;
slot.num_prompt_tokens_processed = slot.num_prompt_tokens;
}
else
{
// push the prompt into the sampling context (do not apply grammar)
for (auto &token : prompt_tokens)
{
llama_sampling_accept(slot.ctx_sampling, ctx, token, false);
}
slot.n_past = common_part(slot.cache_tokens, prompt_tokens);
// the last token of the cache is not in the KV cache until the next call to llama_decode
// (it was sampled, pushed into the "cache_tokens", but not yet put in the context)
if (slot.n_past > 0 && slot.n_past == (int32_t) slot.cache_tokens.size())
{
slot.n_past -= 1;
}
slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past;
if (slot.ga_n != 1)
{
int ga_i = 0;
int32_t ga_n = slot.ga_n;
int32_t ga_w = slot.ga_w;
int32_t slot_npast = 0;
for (int k = 0; k < slot.n_past; ++k)
{
while (slot_npast >= ga_i + ga_w) {
const int bd = (ga_w/ga_n)*(ga_n - 1);
slot_npast -= bd;
ga_i += ga_w/ga_n;
}
slot_npast++;
}
slot.n_past_se = slot_npast;
slot.ga_i = ga_i;
}
LOG_INFO("slot progression", {
{ "slot_id", slot.id },
{ "task_id", slot.task_id },
{ "n_past", slot.n_past },
{ "num_prompt_tokens_processed", slot.num_prompt_tokens_processed }
});
}
slot.cache_tokens = prompt_tokens;
if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
{
// we have to evaluate at least 1 token to generate logits.
LOG_INFO("we have to evaluate at least 1 token to generate logits", {
{ "slot_id", slot.id },
{ "task_id", slot.task_id }
});
slot.n_past--;
if (slot.ga_i > 0)
{
slot.n_past_se--;
}
}
int p0 = (int) system_tokens.size() + slot.n_past;
LOG_INFO("kv cache rm [p0, end)", {
{ "slot_id", slot.id },
{ "task_id", slot.task_id },
{ "p0", p0 }
});
llama_kv_cache_seq_rm(ctx, slot.id, p0, -1);
LOG_VERBOSE("prompt ingested", {
{"n_past", slot.n_past},
{"cached", tokens_to_str(ctx, slot.cache_tokens.cbegin(), slot.cache_tokens.cbegin() + slot.n_past)},
{"to_eval", tokens_to_str(ctx, slot.cache_tokens.cbegin() + slot.n_past, slot.cache_tokens.cend())},
});
const bool has_images = process_images(slot);
// process the prefix of first image
std::vector<llama_token> prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, add_bos_token) : prompt_tokens;
int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past;
int32_t ga_i = slot.ga_i;
int32_t ga_n = slot.ga_n;
int32_t ga_w = slot.ga_w;
for (; slot.n_past < (int) prefix_tokens.size(); ++slot.n_past)
{
if (slot.ga_n != 1)
{
while (slot_npast >= ga_i + ga_w) {
const int bd = (ga_w/ga_n)*(ga_n - 1);
slot_npast -= bd;
ga_i += ga_w/ga_n;
}
}
llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot_npast, {slot.id }, false);
slot_npast++;
}
if (has_images && !ingest_images(slot, n_batch))
{
LOG_ERROR("failed processing images", {
"slot_id", slot.id,
"task_id", slot.task_id,
});
// FIXME @phymbert: to be properly tested
// early returning without changing the slot state will block the slot for ever
// no one at the moment is checking the return value
return false;
}
// extract the logits only for the last token
if (batch.n_tokens > 0)
{
batch.logits[batch.n_tokens - 1] = true;
}
slot.n_decoded = 0;
slot.i_batch = batch.n_tokens - 1;
}
}
}
if (batch.n_tokens == 0)
{
all_slots_are_idle = true;
return true;
}
for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch)
{
const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
for (auto & slot : slots)
{
if (slot.ga_n != 1)
{
// context extension via Self-Extend
while (slot.n_past_se >= slot.ga_i + slot.ga_w)
{
const int ib = (slot.ga_n * slot.ga_i) / slot.ga_w;
const int bd = (slot.ga_w / slot.ga_n) * (slot.ga_n - 1);
const int dd = (slot.ga_w / slot.ga_n) - ib * bd - slot.ga_w;
LOG_TEE("\n");
LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i, slot.n_past_se, ib * bd, slot.ga_i + ib * bd, slot.n_past_se + ib * bd);
LOG_TEE("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n, (slot.ga_i + ib * bd) / slot.ga_n, (slot.ga_i + ib * bd + slot.ga_w) / slot.ga_n);
LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd, slot.ga_i + ib * bd + slot.ga_w + dd, slot.n_past_se + ib * bd + dd);
llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd);
llama_kv_cache_seq_div(ctx, slot.id, slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w,slot.ga_n);
llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd);
slot.n_past_se -= bd;
slot.ga_i += slot.ga_w / slot.ga_n;
LOG_TEE("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", slot.n_past_se + bd, slot.n_past_se, slot.ga_i);
}
slot.n_past_se += n_tokens;
}
}
llama_batch batch_view =
{
n_tokens,
batch.token + i,
nullptr,
batch.pos + i,
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
const int ret = llama_decode(ctx, batch_view);
if (ret != 0)
{
if (n_batch == 1 || ret < 0)
{
// if you get here, it means the KV cache is full - try increasing it via the context size
LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
return false;
}
LOG_TEE("%s : failed to find free space in the KV cache, retrying with smaller n_batch = %d\n", __func__, n_batch / 2);
// retry with half the batch size to try to find a free slot in the KV cache
n_batch /= 2;
i -= n_batch;
continue;
}
for (auto & slot : slots)
{
if (slot.i_batch < (int) i || slot.i_batch >= (int) (i + n_tokens))
{
continue;
}
// prompt evaluated for embedding
if (slot.embedding)
{
send_embedding(slot);
slot.release();
slot.i_batch = -1;
continue;
}
completion_token_output result;
const llama_token id = llama_sampling_sample(slot.ctx_sampling, ctx, NULL, slot.i_batch - i);
llama_sampling_accept(slot.ctx_sampling, ctx, id, true);
slot.n_decoded += 1;
if (slot.n_decoded == 1)
{
slot.t_start_genereration = ggml_time_us();
slot.t_prompt_processing = (slot.t_start_genereration - slot.t_start_process_prompt) / 1e3;
metrics.on_prompt_eval(slot);
}
llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false };
result.tok = id;
const int32_t n_probs = slot.sparams.n_probs;
if (slot.sparams.temp <= 0 && n_probs > 0)
{
// for llama_sample_token_greedy we need to sort candidates
llama_sample_softmax(ctx, &cur_p);
}
for (size_t i = 0; i < std::min(cur_p.size, (size_t)n_probs); ++i)
{
result.probs.push_back({cur_p.data[i].id, cur_p.data[i].p});
}
if (!process_token(result, slot))
{
slot.release();
slot.print_timings();
send_final_response(slot);
metrics.on_prediction(slot);
}
slot.i_batch = -1;
}
}
LOG_VERBOSE("slots updated", {});
return true;
}
void run_on_all_tasks_finished() {
update_slots();
}
};
/* llama.cpp completion api semantics */
static json format_partial_response(
llama_server_context &llama, llama_client_slot *slot, const std::string &content, const std::vector<completion_token_output> &probs
) {
json res = json
{
{"content", content },
{"stop", false},
{"slot_id", slot->id },
{"multimodal", llama.multimodal }
};
if (slot->sparams.n_probs > 0)
{
res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs);
}
return res;
}
struct token_translator
{
llama_context * ctx;
std::string operator()(llama_token tok) const { return llama_token_to_piece(ctx, tok); }
std::string operator()(const completion_token_output &cto) const { return (*this)(cto.tok); }
};
static void append_to_generated_text_from_generated_token_probs(llama_server_context &llama, llama_client_slot *slot)
{
auto & gtps = slot->generated_token_probs;
auto translator = token_translator{llama.ctx};
auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); };
const size_t len = std::accumulate(gtps.begin(), gtps.end(), size_t(0), add_strlen);
if (slot->generated_text.capacity() < slot->generated_text.size() + len)
{
slot->generated_text.reserve(slot->generated_text.size() + len);
}
for (const completion_token_output & cto : gtps)
{
slot->generated_text += translator(cto);
}
}
std::function<void(int)> shutdown_handler;
inline void signal_handler(int signal) { shutdown_handler(signal); }
/////////////////////////////////
////////////////////////////////
//////// LOCALAI code starts below here
/////////////////////////////////
////////////////////////////////
bool loaded_model; // TODO: add a mutex for this, but happens only once loading the model
// The class has a llama instance that is shared across all RPCs
llama_server_context llama;
static void start_llama_server() {
// Wait for model to be loaded first
while (!loaded_model) {
std::this_thread::sleep_for(std::chrono::milliseconds(100));
}
llama.queue_tasks.on_new_task(std::bind(
&llama_server_context::process_single_task, &llama, std::placeholders::_1));
llama.queue_tasks.on_finish_multitask(std::bind(
&llama_server_context::on_finish_multitask, &llama, std::placeholders::_1));
llama.queue_tasks.on_all_tasks_finished(std::bind(
&llama_server_context::run_on_all_tasks_finished, &llama));
llama.queue_results.on_multitask_update(std::bind(
&llama_server_queue::update_multitask,
&llama.queue_tasks,
std::placeholders::_1,
std::placeholders::_2,
std::placeholders::_3
));
llama.queue_tasks.start_loop();
}
json parse_options(bool streaming, const backend::PredictOptions* predict, llama_server_context &llama)
{
// This is for example a slot data from the json data
// slot->params.stream = json_value(data, "stream", false);
// slot->params.cache_prompt = json_value(data, "cache_prompt", false);
// slot->params.n_predict = json_value(data, "n_predict", default_params.n_predict);
// slot->sparams.top_k = json_value(data, "top_k", default_sparams.top_k);
// slot->sparams.top_p = json_value(data, "top_p", default_sparams.top_p);
// slot->sparams.tfs_z = json_value(data, "tfs_z", default_sparams.tfs_z);
// slot->sparams.typical_p = json_value(data, "typical_p", default_sparams.typical_p);
// slot->sparams.temp = json_value(data, "temperature", default_sparams.temp);
// slot->sparams.penalty_last_n = json_value(data, "repeat_last_n", default_sparams.penalty_last_n);
// slot->sparams.penalty_repeat = json_value(data, "repeat_penalty", default_sparams.penalty_repeat);
// slot->sparams.penalty_freq = json_value(data, "frequency_penalty", default_sparams.penalty_freq);
// slot->sparams.penalty_present = json_value(data, "presence_penalty", default_sparams.penalty_present);
// slot->sparams.mirostat = json_value(data, "mirostat", default_sparams.mirostat);
// slot->sparams.mirostat_tau = json_value(data, "mirostat_tau", default_sparams.mirostat_tau);
// slot->sparams.mirostat_eta = json_value(data, "mirostat_eta", default_sparams.mirostat_eta);
// slot->sparams.penalize_nl = json_value(data, "penalize_nl", default_sparams.penalize_nl);
// slot->params.n_keep = json_value(data, "n_keep", slot->params.n_keep);
// slot->params.seed = json_value(data, "seed", default_params.seed);
// slot->sparams.grammar = json_value(data, "grammar", default_sparams.grammar);
// slot->sparams.n_probs = json_value(data, "n_probs", default_sparams.n_probs);
// Create now a json data from the prediction options instead
//
json data;
data["stream"] = streaming;
data["cache_prompt"] = predict->promptcacheall();
data["n_predict"] = predict->tokens() == 0 ? -1 : predict->tokens();
data["top_k"] = predict->topk();
data["top_p"] = predict->topp();
data["tfs_z"] = predict->tailfreesamplingz();
data["typical_p"] = predict->typicalp();
data["temperature"] = predict->temperature();
data["repeat_last_n"] = predict->repeat();
data["repeat_penalty"] = predict->penalty();
data["frequency_penalty"] = predict->frequencypenalty();
data["presence_penalty"] = predict->presencepenalty();
data["mirostat"] = predict->mirostat();
data["mirostat_tau"] = predict->mirostattau();
data["mirostat_eta"] = predict->mirostateta();
data["penalize_nl"] = predict->penalizenl();
data["n_keep"] = predict->nkeep();
data["seed"] = predict->seed();
data["grammar"] = predict->grammar();
data["prompt"] = predict->prompt();
data["ignore_eos"] = predict->ignoreeos();
data["embeddings"] = predict->embeddings();
// for each image in the request, add the image data
//
for (int i = 0; i < predict->images_size(); i++) {
data["image_data"].push_back(json
{
{"id", i},
{"data", predict->images(i)},
});
}
data["stop"] = predict->stopprompts();
// data["n_probs"] = predict->nprobs();
//TODO: images,
return data;
}
// static void parse_options_completion(bool streaming,const backend::PredictOptions* predict, llama_server_context &llama)
// {
// // https://github.com/ggerganov/llama.cpp/blob/d9b33fe95bd257b36c84ee5769cc048230067d6f/examples/server/server.cpp#L673
// gpt_params default_params;
// llama.stream = streaming;
// llama.params.n_predict = predict->tokens() == 0 ? -1 : predict->tokens();
// llama.params.sparams.top_k = predict->topk();
// llama.params.sparams.top_p = predict->topp();
// llama.params.sparams.tfs_z = predict->tailfreesamplingz();
// llama.params.sparams.typical_p = predict->typicalp();
// llama.params.sparams.penalty_last_n = predict->repeat();
// llama.params.sparams.temp = predict->temperature();
// llama.params.sparams.penalty_repeat = predict->penalty();
// llama.params.sparams.penalty_present = predict->presencepenalty();
// llama.params.sparams.penalty_freq = predict->frequencypenalty();
// llama.params.sparams.mirostat = predict->mirostat();
// llama.params.sparams.mirostat_tau = predict->mirostattau();
// llama.params.sparams.mirostat_eta = predict->mirostateta();
// llama.params.sparams.penalize_nl = predict->penalizenl();
// llama.params.n_keep = predict->nkeep();
// llama.params.seed = predict->seed();
// llama.params.sparams.grammar = predict->grammar();
// // llama.params.n_probs = predict->
// llama.params.prompt = predict->prompt();
// llama.params.sparams.logit_bias.clear();
// if (predict->ignoreeos())
// {
// llama.params.sparams.logit_bias[llama_token_eos(llama.model)] = -INFINITY;
// }
// // const auto &logit_bias = body.find("logit_bias");
// // if (logit_bias != body.end() && logit_bias->is_array())
// // {
// // const int n_vocab = llama_n_vocab(llama.model);
// // for (const auto &el : *logit_bias)
// // {
// // if (el.is_array() && el.size() == 2 && el[0].is_number_integer())
// // {
// // llama_token tok = el[0].get<llama_token>();
// // if (tok >= 0 && tok < n_vocab)
// // {
// // if (el[1].is_number())
// // {
// // llama.params.logit_bias[tok] = el[1].get<float>();
// // }
// // else if (el[1].is_boolean() && !el[1].get<bool>())
// // {
// // llama.params.logit_bias[tok] = -INFINITY;
// // }
// // }
// // }
// // }
// // }
// llama.params.antiprompt.clear();
// for (const std::string& stopPrompt : predict->stopprompts()) {
// if (!stopPrompt.empty())
// {
// llama.params.antiprompt.push_back(stopPrompt);
// }
// }
// }
static void params_parse(const backend::ModelOptions* request,
gpt_params & params) {
// this is comparable to: https://github.com/ggerganov/llama.cpp/blob/d9b33fe95bd257b36c84ee5769cc048230067d6f/examples/server/server.cpp#L1809
params.model = request->modelfile();
if (!request->mmproj().empty()) {
// get the directory of modelfile
std::string model_dir = params.model.substr(0, params.model.find_last_of("/\\"));
params.mmproj = model_dir + "/"+ request->mmproj();
}
// params.model_alias ??
params.model_alias = request->modelfile();
params.n_ctx = request->contextsize();
//params.memory_f16 = request->f16memory();
params.n_threads = request->threads();
params.n_gpu_layers = request->ngpulayers();
params.n_batch = request->nbatch();
// Set params.n_parallel by environment variable (LLAMA_PARALLEL), defaults to 1
//params.n_parallel = 1;
const char *env_parallel = std::getenv("LLAMACPP_PARALLEL");
if (env_parallel != NULL) {
params.n_parallel = std::stoi(env_parallel);
params.cont_batching = true;
} else {
params.n_parallel = 1;
}
const char *llama_grpc_servers = std::getenv("LLAMACPP_GRPC_SERVERS");
if (llama_grpc_servers != NULL) {
params.rpc_servers = std::string(llama_grpc_servers);
}
// TODO: Add yarn
if (!request->tensorsplit().empty()) {
std::string arg_next = request->tensorsplit();
// split string by , and /
const std::regex regex{ R"([,/]+)" };
std::sregex_token_iterator it{ arg_next.begin(), arg_next.end(), regex, -1 };
std::vector<std::string> split_arg{ it, {} };
GGML_ASSERT(split_arg.size() <= llama_max_devices());
for (size_t i_device = 0; i_device < llama_max_devices(); ++i_device) {
if (i_device < split_arg.size()) {
params.tensor_split[i_device] = std::stof(split_arg[i_device]);
}
else {
params.tensor_split[i_device] = 0.0f;
}
}
}
if (!request->maingpu().empty()) {
params.main_gpu = std::stoi(request->maingpu());
}
if (!request->loraadapter().empty() && !request->lorabase().empty()) {
float scale_factor = 1.0f;
if (request->lorascale() != 0.0f) {
scale_factor = request->lorascale();
}
// get the directory of modelfile
std::string model_dir = params.model.substr(0, params.model.find_last_of("/\\"));
params.lora_adapters.push_back({ model_dir + "/"+request->loraadapter(), scale_factor });
}
params.use_mlock = request->mlock();
params.use_mmap = request->mmap();
params.flash_attn = request->flashattention();
params.no_kv_offload = request->nokvoffload();
params.embedding = request->embeddings();
if (request->ropescaling() == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
else if (request->ropescaling() == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
else { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
if ( request->yarnextfactor() != 0.0f ) {
params.yarn_ext_factor = request->yarnextfactor();
}
if ( request->yarnattnfactor() != 0.0f ) {
params.yarn_attn_factor = request->yarnattnfactor();
}
if ( request->yarnbetafast() != 0.0f ) {
params.yarn_beta_fast = request->yarnbetafast();
}
if ( request->yarnbetaslow() != 0.0f ) {
params.yarn_beta_slow = request->yarnbetaslow();
}
if ( request->ropefreqbase() != 0.0f ) {
params.rope_freq_base = request->ropefreqbase();
}
if ( request->ropefreqscale() != 0.0f ) {
params.rope_freq_scale = request->ropefreqscale();
}
}
// GRPC Server start
class BackendServiceImpl final : public backend::Backend::Service {
public:
grpc::Status Health(ServerContext* context, const backend::HealthMessage* request, backend::Reply* reply) {
// Implement Health RPC
reply->set_message("OK");
return Status::OK;
}
grpc::Status LoadModel(ServerContext* context, const backend::ModelOptions* request, backend::Result* result) {
// Implement LoadModel RPC
gpt_params params;
params_parse(request, params);
llama_backend_init();
llama_numa_init(params.numa);
// load the model
if (!llama.load_model(params))
{
result->set_message("Failed loading model");
result->set_success(false);
return Status::CANCELLED;
}
llama.initialize();
result->set_message("Loading succeeded");
result->set_success(true);
loaded_model = true;
return Status::OK;
}
grpc::Status PredictStream(grpc::ServerContext* context, const backend::PredictOptions* request, grpc::ServerWriter<backend::Reply>* writer) override {
json data = parse_options(true, request, llama);
const int task_id = llama.queue_tasks.get_new_id();
llama.queue_results.add_waiting_task_id(task_id);
llama.request_completion(task_id, data, false, false, -1);
while (true)
{
task_result result = llama.queue_results.recv(task_id);
if (!result.error) {
const std::string str =
"data: " +
result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) +
"\n\n";
LOG_VERBOSE("data stream", {
{ "to_send", str }
});
backend::Reply reply;
// print it
std::string completion_text = result.result_json.value("content", "");
reply.set_message(completion_text);
int32_t tokens_predicted = result.result_json.value("tokens_predicted", 0);
reply.set_tokens(tokens_predicted);
int32_t tokens_evaluated = result.result_json.value("tokens_evaluated", 0);
reply.set_prompt_tokens(tokens_evaluated);
// Send the reply
writer->Write(reply);
if (result.stop) {
break;
}
} else {
break;
}
}
return grpc::Status::OK;
}
grpc::Status Predict(ServerContext* context, const backend::PredictOptions* request, backend::Reply* reply) {
json data = parse_options(false, request, llama);
const int task_id = llama.queue_tasks.get_new_id();
llama.queue_results.add_waiting_task_id(task_id);
llama.request_completion(task_id, data, false, false, -1);
std::string completion_text;
task_result result = llama.queue_results.recv(task_id);
if (!result.error && result.stop) {
completion_text = result.result_json.value("content", "");
int32_t tokens_predicted = result.result_json.value("tokens_predicted", 0);
int32_t tokens_evaluated = result.result_json.value("tokens_evaluated", 0);
reply->set_prompt_tokens(tokens_evaluated);
reply->set_tokens(tokens_predicted);
reply->set_message(completion_text);
}
else
{
return grpc::Status::OK;
}
return grpc::Status::OK;
}
/// https://github.com/ggerganov/llama.cpp/blob/aa2341298924ac89778252015efcb792f2df1e20/examples/server/server.cpp#L2969
grpc::Status Embedding(ServerContext* context, const backend::PredictOptions* request, backend::EmbeddingResult* embeddingResult) {
json data = parse_options(false, request, llama);
const int task_id = llama.queue_tasks.get_new_id();
llama.queue_results.add_waiting_task_id(task_id);
llama.request_completion(task_id, { {"prompt", data["embeddings"]}, { "n_predict", 0}, {"image_data", ""} }, false, true, -1);
// get the result
task_result result = llama.queue_results.recv(task_id);
//std::cout << "Embedding result JSON" << result.result_json.dump() << std::endl;
llama.queue_results.remove_waiting_task_id(task_id);
if (!result.error && result.stop) {
std::vector<float> embeddings = result.result_json.value("embedding", std::vector<float>());
// loop the vector and set the embeddings results
for (int i = 0; i < embeddings.size(); i++) {
embeddingResult->add_embeddings(embeddings[i]);
}
}
else
{
return grpc::Status::OK;
}
return grpc::Status::OK;
}
};
void RunServer(const std::string& server_address) {
BackendServiceImpl service;
ServerBuilder builder;
builder.AddListeningPort(server_address, grpc::InsecureServerCredentials());
builder.RegisterService(&service);
std::unique_ptr<Server> server(builder.BuildAndStart());
std::cout << "Server listening on " << server_address << std::endl;
server->Wait();
}
int main(int argc, char** argv) {
std::string server_address("localhost:50051");
// Define long and short options
struct option long_options[] = {
{"addr", required_argument, nullptr, 'a'},
{nullptr, 0, nullptr, 0}
};
// Parse command-line arguments
int option;
int option_index = 0;
while ((option = getopt_long(argc, argv, "a:", long_options, &option_index)) != -1) {
switch (option) {
case 'a':
server_address = optarg;
break;
default:
std::cerr << "Usage: " << argv[0] << " [--addr=<address>] or [-a <address>]" << std::endl;
return 1;
}
}
// run the HTTP server in a thread - see comment below
std::thread t([&]()
{
RunServer(server_address);
return 0;
});
//);
start_llama_server();
std::cout << "stopping" << std::endl;
t.join();
llama_backend_free();
return 0;
}
|