File size: 97,669 Bytes
1d08aca b5f596d 761e1b2 b5f596d 1d08aca 761e1b2 1d08aca 761e1b2 1d08aca 761e1b2 1d08aca 761e1b2 1d08aca 761e1b2 1d08aca 761e1b2 1d08aca 761e1b2 1d08aca 761e1b2 1d08aca 761e1b2 1d08aca b5f596d 1d08aca | 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 | import os
if os.environ.get("MODELSCOPE_ENVIRONMENT") == "studio":
from modelscope import patch_hub
patch_hub()
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:256"
from config import CONFIG, ModelConfig
from utils import (
cleanMessages,
parse_think_response,
remove_nested_think_tags_stack,
format_bytes,
log,
detect_tools_and_reasoning,
universal_tool,
)
import copy, types, gc, sys, re, time, collections, asyncio
from huggingface_hub import hf_hub_download
from loguru import logger
from rich import print
from snowflake import SnowflakeGenerator
CompletionIdGenerator = SnowflakeGenerator(42, timestamp=1741101491595)
from typing import List, Optional, Union, Any, Dict
import uuid
from pydantic import BaseModel, Field, model_validator
from pydantic_settings import BaseSettings
import numpy as np
import torch
if "cuda" in CONFIG.STRATEGY.lower() and not torch.cuda.is_available():
logger.info(f"CUDA not found, fall back to cpu")
CONFIG.STRATEGY = "cpu fp16"
# Normalize STRATEGY to include precision if missing (e.g., 'cpu' -> 'cpu fp16')
_s = CONFIG.STRATEGY.lower()
if ("cpu" in _s or "cuda" in _s) and not ("fp16" in _s or "fp32" in _s):
logger.info(f"STRATEGY missing precision, appending 'fp16' to `{CONFIG.STRATEGY}`")
CONFIG.STRATEGY = CONFIG.STRATEGY + " fp16"
try:
from pynvml import nvmlInit, nvmlDeviceGetHandleByIndex, nvmlDeviceGetMemoryInfo
except Exception:
nvmlInit = None
nvmlDeviceGetHandleByIndex = None
nvmlDeviceGetMemoryInfo = None
if "cuda" in CONFIG.STRATEGY.lower() and nvmlInit is not None and nvmlDeviceGetHandleByIndex is not None:
nvmlInit()
gpu_h = nvmlDeviceGetHandleByIndex(0)
def logGPUState():
if "cuda" in CONFIG.STRATEGY and nvmlDeviceGetMemoryInfo is not None:
gpu_info = nvmlDeviceGetMemoryInfo(gpu_h)
logger.info(
f"[STATUS] Torch - {format_bytes(torch.cuda.memory_allocated())} - NVML - vram {format_bytes(gpu_info.total)} used {format_bytes(gpu_info.used)} free {format_bytes(gpu_info.free)}"
)
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.allow_tf32 = True
torch.backends.cuda.matmul.allow_tf32 = True
os.environ["RWKV_V7_ON"] = "1" # enable this for rwkv-7 models
os.environ["RWKV_JIT_ON"] = "1"
os.environ["RWKV_CUDA_ON"] = (
"1" if CONFIG.RWKV_CUDA_ON and "cuda" in CONFIG.STRATEGY.lower() else "0"
)
from rwkv.model import RWKV
from rwkv.utils import PIPELINE, PIPELINE_ARGS
from fastapi import FastAPI, HTTPException, UploadFile, File
from starlette.background import BackgroundTask
from fastapi.responses import StreamingResponse
from fastapi.middleware.cors import CORSMiddleware
from fastapi.staticfiles import StaticFiles
from fastapi.middleware.gzip import GZipMiddleware
from api_types import (
ChatMessage,
ChatCompletion,
ChatCompletionChunk,
Usage,
PromptTokensDetails,
ChatCompletionChoice,
ChatCompletionMessage,
SamplerConfig,
UploadedFile,
FileUploadResponse,
)
class ModelStorage:
MODEL_CONFIG: Optional[ModelConfig] = None
model: Optional[RWKV] = None
pipeline: Optional[PIPELINE] = None
MODEL_STORAGE: Dict[str, ModelStorage] = {}
DEFALUT_MODEL_NAME = None
DEFAULT_REASONING_MODEL_NAME = None
# In-memory model state store to support streaming continuation/resume per state_name.
# Keys: (model_name, state_name) -> dict with 'state' and 'model_tokens'
STATE_STORE: Dict[tuple, Any] = {}
# Serialized state store file path and flush interval defined in CONFIG
_STATE_STORE_PATH = getattr(CONFIG, 'STATE_STORE_PATH', './state_store.json')
_LAST_STATE_STORE_WRITE = 0
# sentinel for model-initiated tool calls: <tool-call>{json}</tool-call>
TOOL_CALL_RE = re.compile(r"<tool-call>\s*(\{.*?\})\s*</tool-call>", re.S)
# File uploads: simple in-memory index (persisted on disk via the files themselves)
UPLOADED_FILES: Dict[str, dict] = {}
def _serialize_state_store() -> dict:
# Save only model_tokens to disk; model_state (torch objects) are not serializable
serial = {}
for (model_name, state_name), entry in STATE_STORE.items():
try:
mt = entry.get('model_tokens') if isinstance(entry, dict) else None
if mt is None:
# if entry is a raw model_state, skip
continue
serial[f"{model_name}|{state_name}"] = {
'model': model_name,
'state_name': state_name,
'model_tokens': mt,
}
except Exception:
continue
return serial
def _load_state_store_from_disk():
global STATE_STORE
try:
if os.path.exists(_STATE_STORE_PATH):
import json
with open(_STATE_STORE_PATH, 'r', encoding='utf-8') as f:
data = json.load(f)
for k, v in data.items():
model = v.get('model')
state_name = v.get('state_name')
model_tokens = v.get('model_tokens')
if model and state_name and isinstance(model_tokens, list):
STATE_STORE[(model, state_name)] = {
'state': None,
'model_tokens': model_tokens,
}
logger.info(f"Loaded {len(STATE_STORE)} entries from state store file {_STATE_STORE_PATH}")
except Exception as e:
logger.info(f"Failed to load state store from disk: {e}")
def _save_state_store_to_disk(force=False):
global _LAST_STATE_STORE_WRITE
now = time.time()
if not force and now - _LAST_STATE_STORE_WRITE < getattr(CONFIG, 'STATE_STORE_FLUSH_INTERVAL', 5):
return
try:
serial = _serialize_state_store()
if not serial:
return
import json
tmp = _STATE_STORE_PATH + ".tmp"
with open(tmp, 'w', encoding='utf-8') as f:
json.dump(serial, f)
os.replace(tmp, _STATE_STORE_PATH)
_LAST_STATE_STORE_WRITE = now
except Exception as e:
logger.info(f"Write state store to disk failed: {e}")
def _recompute_out_and_state_from_tokens(model_name: str, model_tokens: List[int]):
"""
Recompute the `out` logits and `model_state` by forwarding through tokens in chunks.
Returns a tuple (out, model_state).
"""
ms = MODEL_STORAGE.get(model_name)
if not ms or not ms.model:
return None, None
model_state = None
out = None
tokens = list(model_tokens) if isinstance(model_tokens, list) else [0]
while len(tokens) > 0:
out, model_state = ms.model.forward(tokens[: CONFIG.CHUNK_LEN], model_state)
tokens = tokens[CONFIG.CHUNK_LEN :]
return out, model_state
def resolve_request_flags(request, detection):
"""Resolve effective booleans for web_search, file_tool, tools, and reasoning
based on request flags (explicit), sampler overrides, model defaults and detection.
Returns dict with keys: web_search_enabled, file_tool_enabled, tools_enabled, reasoning_enabled.
"""
# Web search
web_search_enabled = (
True
if (request.enable_web_search is not None and request.enable_web_search)
else (
request.web_search
or (request.auto_web_search if request.auto_web_search is not None else (getattr(CONFIG, 'AUTO_ENABLE_WEB_SEARCH', True) and detection.get('need_web_search')))
)
)
if not getattr(CONFIG, 'ENABLE_WEB_SEARCH_BY_DEFAULT', True) and request.enable_web_search is None and not (request.web_search or False):
web_search_enabled = False
try:
if request.sampler and getattr(request.sampler, 'ALLOW_WEB_SEARCH', None) is not None:
web_search_enabled = bool(request.sampler.ALLOW_WEB_SEARCH)
elif hasattr(request, 'sampler_allow_web_search') and request.sampler_allow_web_search is not None:
web_search_enabled = bool(request.sampler_allow_web_search)
else:
ms = MODEL_STORAGE.get(request.model)
if ms and ms.MODEL_CONFIG:
if hasattr(ms.MODEL_CONFIG, 'DEFAULT_SAMPLER') and getattr(ms.MODEL_CONFIG.DEFAULT_SAMPLER, 'ALLOW_WEB_SEARCH', None) is not None:
web_search_enabled = bool(ms.MODEL_CONFIG.DEFAULT_SAMPLER.ALLOW_WEB_SEARCH)
elif hasattr(ms.MODEL_CONFIG, 'ALLOW_WEB_SEARCH') and not ms.MODEL_CONFIG.ALLOW_WEB_SEARCH:
web_search_enabled = False
except Exception:
pass
# File tool decision
if request.enable_file_tool is not None:
file_tool_enabled = bool(request.enable_file_tool)
else:
auto_file_flag = request.auto_file_tool if request.auto_file_tool is not None else getattr(CONFIG, 'AUTO_ENABLE_TOOLS', True)
file_tool_enabled = bool((request.file_ids and len(request.file_ids) > 0) or (auto_file_flag and request.file_ids))
if not getattr(CONFIG, 'ALLOW_FILE_TOOL_BY_DEFAULT', True) and request.enable_file_tool is None:
file_tool_enabled = False
try:
if request.sampler and getattr(request.sampler, 'ALLOW_FILE_TOOL', None) is not None:
file_tool_enabled = bool(request.sampler.ALLOW_FILE_TOOL)
elif hasattr(request, 'sampler_allow_file_tool') and request.sampler_allow_file_tool is not None:
file_tool_enabled = bool(request.sampler_allow_file_tool)
else:
ms = MODEL_STORAGE.get(request.model)
if ms and ms.MODEL_CONFIG:
if hasattr(ms.MODEL_CONFIG, 'DEFAULT_SAMPLER') and getattr(ms.MODEL_CONFIG.DEFAULT_SAMPLER, 'ALLOW_FILE_TOOL', None) is not None:
file_tool_enabled = bool(ms.MODEL_CONFIG.DEFAULT_SAMPLER.ALLOW_FILE_TOOL)
elif hasattr(ms.MODEL_CONFIG, 'ALLOW_FILE_TOOL') and not ms.MODEL_CONFIG.ALLOW_FILE_TOOL:
file_tool_enabled = False
except Exception:
pass
# Tools decision
if request.enable_tools is not None:
tools_enabled = bool(request.enable_tools)
else:
auto_tools_flag = request.auto_tools if request.auto_tools is not None else getattr(CONFIG, 'AUTO_ENABLE_TOOLS', True)
tools_enabled = bool(request.tools) or getattr(CONFIG, 'ENABLE_TOOLS_BY_DEFAULT', False) or (auto_tools_flag and (detection.get('need_calc') or detection.get('need_web_search')))
try:
if request.sampler and getattr(request.sampler, 'ALLOW_TOOLS', None) is not None:
tools_enabled = bool(request.sampler.ALLOW_TOOLS)
elif hasattr(request, 'sampler_allow_tools') and request.sampler_allow_tools is not None:
tools_enabled = bool(request.sampler_allow_tools)
else:
ms = MODEL_STORAGE.get(request.model)
if ms and ms.MODEL_CONFIG:
if hasattr(ms.MODEL_CONFIG, 'DEFAULT_SAMPLER') and getattr(ms.MODEL_CONFIG.DEFAULT_SAMPLER, 'ALLOW_TOOLS', None) is not None:
if not ms.MODEL_CONFIG.DEFAULT_SAMPLER.ALLOW_TOOLS:
tools_enabled = False
elif hasattr(ms.MODEL_CONFIG, 'ALLOW_TOOLS') and not ms.MODEL_CONFIG.ALLOW_TOOLS:
tools_enabled = False
except Exception:
pass
# Reasoning decision
reasoning_enabled = bool(
True
if (request.enable_reasoning is not None and request.enable_reasoning)
else (
bool(False) or bool(request.auto_reasoning if request.auto_reasoning is not None else (getattr(CONFIG, 'AUTO_ENABLE_REASONING', True) and bool(detection.get('need_reasoning'))))
)
)
if not getattr(CONFIG, 'ENABLE_REASONING_BY_DEFAULT', True) and request.enable_reasoning is None:
reasoning_enabled = False
try:
if request.sampler and getattr(request.sampler, 'ALLOW_REASONING', None) is not None:
reasoning_enabled = bool(request.sampler.ALLOW_REASONING)
elif hasattr(request, 'sampler_allow_reasoning') and request.sampler_allow_reasoning is not None:
reasoning_enabled = bool(request.sampler_allow_reasoning)
else:
ms = MODEL_STORAGE.get(request.model)
if ms and ms.MODEL_CONFIG:
if hasattr(ms.MODEL_CONFIG, 'DEFAULT_SAMPLER') and getattr(ms.MODEL_CONFIG.DEFAULT_SAMPLER, 'ALLOW_REASONING', None) is not None:
if not ms.MODEL_CONFIG.DEFAULT_SAMPLER.ALLOW_REASONING:
reasoning_enabled = False
elif hasattr(ms.MODEL_CONFIG, 'ALLOW_REASONING') and not ms.MODEL_CONFIG.ALLOW_REASONING:
reasoning_enabled = False
except Exception:
pass
# Also apply model-level disable
try:
ms = MODEL_STORAGE.get(request.model)
if ms and ms.MODEL_CONFIG and hasattr(ms.MODEL_CONFIG, 'ALLOW_REASONING') and not ms.MODEL_CONFIG.ALLOW_REASONING:
reasoning_enabled = False
except Exception:
pass
return {
'web_search_enabled': web_search_enabled,
'file_tool_enabled': file_tool_enabled,
'tools_enabled': tools_enabled,
'reasoning_enabled': reasoning_enabled,
}
# Move ChatCompletionRequest definition above fallback apply_model_tags_to_request
# Move ChatCompletionRequest definition above fallback apply_model_tags_to_request
try:
from model_tags import apply_model_tags_to_request_obj as apply_model_tags_to_request
except Exception:
def apply_model_tags_to_request(req: Any):
# Fallback implementation if the module import fails; keep behavior robust
if not req or not getattr(req, 'model', None) or ':' not in req.model:
return
original = req.model
parts = [p.strip() for p in original.split(":") if p is not None and p != ""]
if len(parts) <= 1:
return
base = parts[0]
tags = parts[1:]
req.model = base
for tag in tags:
t = tag.lower()
if t in ("thinking", "think", "reasoning", "reason"):
req.enable_reasoning = True
req.auto_reasoning = False
elif t in ("web", "web_search", "search"):
req.enable_web_search = True
req.web_search = True
req.auto_web_search = False
elif t in ("no-web", "disable-web", "no-web-search"):
req.enable_web_search = False
req.web_search = False
elif t in ("tools", "enable-tools"):
req.enable_tools = True
req.auto_tools = False
elif t in ("no-tools", "disable-tools"):
req.enable_tools = False
elif t in ("file", "file_tool", "filetool"):
req.enable_file_tool = True
req.auto_file_tool = False
elif t in ("no-file", "disable-file"):
req.enable_file_tool = False
elif t in ("universal", "univ"):
req.enable_universal = True
req.auto_universal = False
elif t in ("stream",):
req.stream = True
logger.info(f"STRATEGY - {CONFIG.STRATEGY}")
logGPUState()
def load_models_once():
"""Load and initialize configured models into `MODEL_STORAGE`. This is executed once at server startup."""
global DEFALUT_MODEL_NAME, DEFAULT_REASONING_MODEL_NAME
logger.info(f"Configured {len(CONFIG.MODELS)} model(s) in ROOT config")
for model_config in CONFIG.MODELS:
logger.info(f"Load Model - {model_config.SERVICE_NAME}")
if model_config.MODEL_FILE_PATH == None:
model_config.MODEL_FILE_PATH = hf_hub_download(
repo_id=str(model_config.DOWNLOAD_MODEL_REPO_ID),
filename=str(model_config.DOWNLOAD_MODEL_FILE_NAME),
local_dir=str(model_config.DOWNLOAD_MODEL_DIR),
)
logger.info(f"Load Model - Path - {model_config.MODEL_FILE_PATH}")
if model_config.DEFAULT_CHAT:
if DEFALUT_MODEL_NAME != None:
logger.info(
f"Load Model - Replace `DEFALUT_MODEL_NAME` from `{DEFALUT_MODEL_NAME}` to `{model_config.SERVICE_NAME}`"
)
DEFALUT_MODEL_NAME = model_config.SERVICE_NAME
if model_config.DEFAULT_REASONING:
if DEFAULT_REASONING_MODEL_NAME != None:
logger.info(
f"Load Model - Replace `DEFAULT_REASONING_MODEL_NAME` from `{DEFAULT_REASONING_MODEL_NAME}` to `{model_config.SERVICE_NAME}`"
)
DEFAULT_REASONING_MODEL_NAME = model_config.SERVICE_NAME
logger.info(f"Load Model - Loading `{model_config.SERVICE_NAME}`")
print(model_config.DEFAULT_SAMPLER)
MODEL_STORAGE[model_config.SERVICE_NAME] = ModelStorage()
MODEL_STORAGE[model_config.SERVICE_NAME].MODEL_CONFIG = model_config
MODEL_STORAGE[model_config.SERVICE_NAME].model = RWKV(
model=model_config.MODEL_FILE_PATH.replace(".pth", ""),
strategy=CONFIG.STRATEGY,
)
MODEL_STORAGE[model_config.SERVICE_NAME].pipeline = PIPELINE(
MODEL_STORAGE[model_config.SERVICE_NAME].model, model_config.VOCAB
)
if "cuda" in CONFIG.STRATEGY:
torch.cuda.empty_cache()
gc.collect()
logGPUState()
logger.info(f"Load Model - DEFALUT_MODEL_NAME is `{DEFALUT_MODEL_NAME}`")
logger.info(f"Load Model - DEFAULT_REASONING_MODEL_NAME is `{DEFAULT_REASONING_MODEL_NAME}`")
if len(MODEL_STORAGE) == 1:
single_name = list(MODEL_STORAGE.keys())[0]
if DEFALUT_MODEL_NAME != single_name:
DEFALUT_MODEL_NAME = single_name
logger.info(f"Load Model - Only one model present; DEFALUT_MODEL_NAME set to `{DEFALUT_MODEL_NAME}`")
if DEFAULT_REASONING_MODEL_NAME != single_name:
DEFAULT_REASONING_MODEL_NAME = single_name
logger.info(f"Load Model - Only one model present; DEFAULT_REASONING_MODEL_NAME set to `{DEFAULT_REASONING_MODEL_NAME}`")
class ChatCompletionRequest(BaseModel):
model: str = Field(
default="rwkv-latest",
description="Specify the model name. Model tags/suffixes (e.g., ':thinking' or ':web') are not supported — set the corresponding request flags (enable_reasoning, web_search, enable_file_tool) instead.",
)
messages: Optional[List[ChatMessage]] = Field(default=None)
prompt: Optional[str] = Field(default=None)
max_tokens: Optional[int] = Field(default=None)
temperature: Optional[float] = Field(default=None)
top_p: Optional[float] = Field(default=None)
presence_penalty: Optional[float] = Field(default=None)
count_penalty: Optional[float] = Field(default=None)
penalty_decay: Optional[float] = Field(default=None)
stream: Optional[bool] = Field(default=True, description="Whether to stream token-by-token responses. If None, uses CONFIG.DEFAULT_STREAM")
state_name: Optional[str] = Field(default=None)
include_usage: Optional[bool] = Field(default=False)
stop: Optional[list[str]] = Field(["\n\n"])
stop_tokens: Optional[list[int]] = Field([0])
# Note: these defaults are intentionally None so the model may decide
# autonomously whether to use web search based on prompt detection unless
# the client explicitly sets flags. `auto_web_search` will be consulted
# when `enable_web_search` and `web_search` are None.
web_search: Optional[bool] = Field(default=None, description="Whether to perform a web search and append results to the prompt; if None, auto-detection is used")
enable_web_search: Optional[bool] = Field(default=None, description="Explicitly enable web search (overrides auto/web_search) if set; if None, auto-detection controls it")
auto_web_search: Optional[bool] = Field(default=True, description="Whether to enable web_search based on auto-detected intent")
enable_tools: Optional[bool] = Field(default=None, description="Explicitly enable tools (overrides auto detection)")
auto_tools: Optional[bool] = Field(default=True, description="Whether to enable tools based on auto-detected intent")
enable_reasoning: Optional[bool] = Field(default=None, description="Explicitly override reasoning enablement; if None, auto-detection controls it")
auto_reasoning: Optional[bool] = Field(default=True, description="Whether to enable reasoning based on auto detection")
enable_universal: Optional[bool] = Field(default=None, description="Explicitly enable the universal tool execution")
auto_universal: Optional[bool] = Field(default=True, description="Whether to auto enable universal tool execution")
search_top_k: Optional[int] = Field(default=3, description="Number of web search results to retrieve")
tools: Optional[List[Dict[str, Any]]] = Field(default=None, description="List of tools to execute server-side (e.g., {'name':'web_search','args':{'query':'x'}})")
# Per-request sampler overrides for ALLOW_* flags. These let the user
# disable server-side features for this particular request if needed.
sampler_allow_web_search: Optional[bool] = Field(default=None, description="Per-request (sampler) override allowing web_search")
sampler_allow_tools: Optional[bool] = Field(default=None, description="Per-request (sampler) override allowing tools")
sampler_allow_reasoning: Optional[bool] = Field(default=None, description="Per-request (sampler) override allowing reasoning")
# Per-request sampler config object; if provided, these settings will
# override the model defaults for this request.
sampler: Optional[SamplerConfig] = Field(default=None, description="Per-request sampler settings (overrides model default)")
# File uploads: allow referencing uploaded files in the request
file_ids: Optional[List[str]] = Field(default=None, description="List of uploaded file IDs that the model may use for this request")
enable_file_tool: Optional[bool] = Field(default=None, description="Explicitly enable file-based tools for this request")
auto_file_tool: Optional[bool] = Field(default=None, description="Auto-detect whether file-based tools are needed")
sampler_allow_file_tool: Optional[bool] = Field(default=None, description="Per-request sampler override allowing file tools")
@model_validator(mode="before")
@classmethod
def validate_mutual_exclusivity(cls, data: Any) -> Any:
if not isinstance(data, dict):
return data
messages_provided = "messages" in data and data["messages"] != None
prompt_provided = "prompt" in data and data["prompt"] != None
if messages_provided and prompt_provided:
raise ValueError("messages and prompt cannot coexist. Choose one.")
if not messages_provided and not prompt_provided:
raise ValueError("Either messages or prompt must be provided.")
return data
app = FastAPI(title="RWKV OpenAI-Compatible API")
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
app.add_middleware(GZipMiddleware, minimum_size=1000, compresslevel=5)
@app.on_event("startup")
async def _startup_state_load_and_persist_loop():
# Load previous persisted state (tokens only) at startup
_load_state_store_from_disk()
# Load configured models once on startup
try:
load_models_once()
except Exception as e:
logger.info(f"Model loading at startup failed: {e}")
async def _persist_loop():
while True:
try:
_save_state_store_to_disk(force=False)
except Exception:
pass
await asyncio.sleep(getattr(CONFIG, 'STATE_STORE_FLUSH_INTERVAL', 5))
# Spawn background flush task
try:
asyncio.create_task(_persist_loop())
except Exception:
pass
async def runPrefill(
request: ChatCompletionRequest, ctx: str, model_tokens: List[int], model_state
):
ctx = ctx.replace("\r\n", "\n")
out = None
ms = MODEL_STORAGE.get(request.model)
if not ms or not ms.pipeline or not ms.model:
raise HTTPException(500, f"Model {request.model} not loaded or pipeline missing")
tokens = ms.pipeline.encode(ctx)
tokens = [int(x) for x in tokens]
model_tokens += tokens
while len(tokens) > 0:
out, model_state = ms.model.forward(
tokens[: CONFIG.CHUNK_LEN], model_state
)
tokens = tokens[CONFIG.CHUNK_LEN :]
await asyncio.sleep(0)
return out, model_tokens, model_state
def generate(
request: ChatCompletionRequest,
out,
model_tokens: List[int],
model_state,
max_tokens=2048,
):
ms = MODEL_STORAGE.get(request.model)
if not ms or not ms.pipeline or not ms.model:
raise HTTPException(500, f"Model {request.model} not loaded or pipeline missing")
# Enforce sampling configuration limits from CONFIG
from config import CONFIG as _CFG
temperature = request.temperature if request.temperature is not None else 0.2
temperature = min(max(temperature, getattr(_CFG, 'MIN_TEMPERATURE', 0.0)), getattr(_CFG, 'MAX_TEMPERATURE', 2.0))
top_p = request.top_p if request.top_p is not None else 0.9
top_p = min(max(top_p, getattr(_CFG, 'MIN_TOP_P', 0.0)), getattr(_CFG, 'MAX_TOP_P', 1.0))
alpha_frequency = request.count_penalty if request.count_penalty is not None else 0.0
alpha_presence = request.presence_penalty if request.presence_penalty is not None else 0.0
penalty_decay = request.penalty_decay if request.penalty_decay is not None else 0.5
args = PIPELINE_ARGS(
temperature=max(0.2, temperature),
top_p=top_p,
alpha_frequency=alpha_frequency,
alpha_presence=alpha_presence,
token_ban=[], # ban the generation of some tokens
token_stop=[0],
) # stop generation whenever you see any token here
occurrence = {}
out_tokens: List[int] = []
out_last = 0
# Stream token-by-token; each chunk contains a single decoded token string.
for i in range(max_tokens):
for n in occurrence:
out[n] -= args.alpha_presence + occurrence[n] * args.alpha_frequency
# out[0] -= 1e10 # disable END_OF_TEXT
token = ms.pipeline.sample_logits(
out, temperature=args.temperature, top_p=args.top_p
)
if token == 0 and request.stop_tokens and token in request.stop_tokens:
yield {
"content": "",
"tokens": out_tokens[out_last:],
"finish_reason": "stop:token:0",
"state": model_state,
}
del out
gc.collect()
return
out, model_state = ms.model.forward([token], model_state)
model_tokens.append(token)
out_tokens.append(token)
if request.stop_tokens and token in request.stop_tokens:
yield {
"content": "",
"tokens": out_tokens[out_last:],
"finish_reason": f"stop:token:{token}",
"state": model_state,
}
del out
gc.collect()
return
for xxx in list(occurrence.keys()):
occurrence[xxx] *= penalty_decay
occurrence[token] = 1 + (occurrence[token] if token in occurrence else 0)
# Decode token to text and yield it as a single-token chunk
decoded = ms.pipeline.decode([token])
# filter out replacement characters
if "\ufffd" in decoded:
continue
yield {
"content": decoded,
"tokens": [token],
"finish_reason": None,
"state": model_state,
}
out_last = i + 1
else:
yield {
"content": "",
"tokens": [],
"finish_reason": "length",
}
async def chatResponse(
request: ChatCompletionRequest,
model_state: Any,
completionId: str,
enableReasoning: bool,
) -> ChatCompletion:
createTimestamp = time.time()
# use global resolve_request_flags
def decide_file_tool_enabled(request):
return resolve_request_flags(request, detection)['file_tool_enabled']
def decide_tools_enabled(request, detection):
return resolve_request_flags(request, detection)['tools_enabled']
def decide_reasoning_enabled(request, detection, enableReasoning):
# resolve_request_flags ignores the enableReasoning baseline, so compute flags then
flags = resolve_request_flags(request, detection)
return flags['reasoning_enabled']
def execute_tools(request, detection, prompt, executed_tool_calls, web_search_enabled, tools_enabled, file_tool_enabled, raw_prompt):
"""Helper to execute tools and update prompt and executed_tool_calls."""
# If file tools are enabled and files are attached, inject them into the prompt (for streaming)
if file_tool_enabled and request.file_ids:
for fid in request.file_ids:
try:
if fid not in UPLOADED_FILES:
continue
meta = UPLOADED_FILES.get(fid)
if not meta:
continue
from utils import file_read_from_path
fpath = meta.get('path')
if not fpath or not os.path.exists(fpath):
continue
file_content = file_read_from_path(fpath, 200000)
if file_content:
exec_entry = {"name": "file_inject", "args": {"file_id": fid}, "result": {"action": "file_inject", "result": "injected", "metadata": {"file_id": fid, "filename": meta.get('filename')}}}
executed_tool_calls.append(exec_entry)
prompt = (f"AttachedFile: {meta.get('filename')} (id:{fid})\n{file_content}\n\n" + prompt)
except Exception as e:
logger.info(f"File injection error: {e}")
# Only a single injection loop above is needed; duplicate removed.
if request.tools:
try:
for tool in request.tools:
name = tool.get('name')
args = tool.get('args', {})
if name == 'web_search':
from utils import web_search
search_q = args.get('query') or (request.prompt if request.prompt else cleanMessages(request.messages or []))
search_top_k = int(args.get('top_k') or request.search_top_k or 3)
search_str = web_search(search_q, search_top_k)
if search_str:
search_res_struct = {"action": "web_search", "result": str(search_str), "metadata": {"query": search_q, "top_k": search_top_k, "confidence": 0.9}}
executed_tool_calls.append({"name": "web_search", "args": {"query": search_q, "top_k": search_top_k}, "result": search_res_struct})
prompt = (f"ToolResults:\n{search_res_struct.get('result')}\n\nUse these results to answer the prompt.\n\n" + prompt)
elif name == 'calc' or name == 'calculator':
from utils import calc
expr = args.get('expression')
if expr:
calc_res = calc(expr)
calc_res_struct = {"action": "calc", "result": str(calc_res), "metadata": {"expression": expr, "confidence": 0.98}}
executed_tool_calls.append({"name": "calc", "args": {"expression": expr}, "result": calc_res_struct})
prompt = (f"ToolResults:\nCalcResult:{expr} = {calc_res_struct.get('result')}\n\nUse this result to answer the prompt.\n\n" + prompt)
elif name == 'universal':
try:
res = universal_tool(args or {"query": raw_prompt}, allow_web_search=bool(web_search_enabled), allow_tools=bool(tools_enabled), allow_file_tool=bool(file_tool_enabled))
if isinstance(res, dict):
result_text = res.get('result') if res.get('result') is not None else ''
else:
result_text = str(res)
executed_tool_calls.append({"name": "universal", "args": args, "result": res})
prompt = (f"ToolResults:\n{result_text}\n\nUse this result to answer the prompt.\n\n" + prompt)
except Exception as e:
logger.info(f"Universal tool execution error: {e}")
else:
logger.info(f"Unsupported tool requested: {name}")
# ...existing code for other tools (fetch_url, summarize, keywords, etc.)...
# For brevity, keep the rest as is, or further refactor if needed.
except Exception as e:
logger.info(f"Tool processing error: {e}")
elif request.web_search or web_search_enabled:
try:
from utils import web_search
search_q = request.prompt if request.prompt else cleanMessages(request.messages or [])
search_res = web_search(search_q, int(request.search_top_k or 3))
if search_res:
search_res_struct = {"action": "web_search", "result": str(search_res), "metadata": {"query": search_q, "top_k": int(request.search_top_k or 3), "confidence": 0.9}}
executed_tool_calls.append({"name": "web_search", "args": {"query": search_q, "top_k": int(request.search_top_k or 3)}, "result": search_res_struct})
prompt = f"WebSearchResults:\n{search_res_struct.get('result')}\n\n" + prompt
except Exception:
pass
return prompt, executed_tool_calls
# Build raw prompt for detection (prefer explicit request.prompt, else messages)
raw_prompt = request.prompt.strip() if request.prompt is not None else cleanMessages(request.messages or [])
detection = detect_tools_and_reasoning(raw_prompt)
prompt = raw_prompt if request.prompt is not None else f"{cleanMessages(request.messages or [])}\n\nAssistant:{' <think' if enableReasoning else ''}"
def decide_web_search_enabled(request, detection):
web_search_enabled = (
True
if (request.enable_web_search is not None and request.enable_web_search)
else (
request.web_search
or (request.auto_web_search if request.auto_web_search is not None else (getattr(CONFIG, 'AUTO_ENABLE_WEB_SEARCH', True) and detection.get('need_web_search')))
)
)
if not getattr(CONFIG, 'ENABLE_WEB_SEARCH_BY_DEFAULT', True) and request.enable_web_search is None and not (request.web_search or False):
web_search_enabled = False
try:
if request.sampler and getattr(request.sampler, 'ALLOW_WEB_SEARCH', None) is not None:
web_search_enabled = bool(request.sampler.ALLOW_WEB_SEARCH)
elif hasattr(request, 'sampler_allow_web_search') and request.sampler_allow_web_search is not None:
web_search_enabled = bool(request.sampler_allow_web_search)
else:
ms = MODEL_STORAGE.get(request.model)
if ms and ms.MODEL_CONFIG:
if hasattr(ms.MODEL_CONFIG, 'DEFAULT_SAMPLER') and getattr(ms.MODEL_CONFIG.DEFAULT_SAMPLER, 'ALLOW_WEB_SEARCH', None) is not None:
web_search_enabled = bool(ms.MODEL_CONFIG.DEFAULT_SAMPLER.ALLOW_WEB_SEARCH)
elif hasattr(ms.MODEL_CONFIG, 'ALLOW_WEB_SEARCH') and not ms.MODEL_CONFIG.ALLOW_WEB_SEARCH:
web_search_enabled = False
except Exception:
pass
return web_search_enabled
web_search_enabled = decide_web_search_enabled(request, detection)
file_tool_enabled = decide_file_tool_enabled(request)
tools_enabled = decide_tools_enabled(request, detection)
reasoning_enabled = decide_reasoning_enabled(request, detection, enableReasoning)
enableReasoning = reasoning_enabled
if request.enable_web_search is None:
request.web_search = web_search_enabled
if tools_enabled and not request.tools:
if detection.get('detected_tools'):
detected = detection.get('detected_tools') or []
filtered = []
for t in detected:
tname = t.get('name')
args = t.get('args', {})
ms_check = MODEL_STORAGE.get(request.model)
allowed = True
try:
if ms_check and ms_check.MODEL_CONFIG:
if tname in ('web_search', 'fetch_url') and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_FETCH_URL') and not ms_check.MODEL_CONFIG.ALLOW_FETCH_URL:
allowed = False
if tname == 'summarize' and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_SUMMARIZE') and not ms_check.MODEL_CONFIG.ALLOW_SUMMARIZE:
allowed = False
if tname in ('keywords',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_KEYWORDS') and not ms_check.MODEL_CONFIG.ALLOW_KEYWORDS:
allowed = False
if tname in ('sentiment',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_SENTIMENT') and not ms_check.MODEL_CONFIG.ALLOW_SENTIMENT:
allowed = False
if tname in ('translate',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_TRANSLATE') and not ms_check.MODEL_CONFIG.ALLOW_TRANSLATE:
allowed = False
if tname in ('spell_check',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_SPELL_CHECK') and not ms_check.MODEL_CONFIG.ALLOW_SPELL_CHECK:
allowed = False
if tname in ('format_code',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_FORMAT_CODE') and not ms_check.MODEL_CONFIG.ALLOW_FORMAT_CODE:
allowed = False
if tname in ('explain_code',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_EXPLAIN_CODE') and not ms_check.MODEL_CONFIG.ALLOW_EXPLAIN_CODE:
allowed = False
except Exception:
pass
if allowed:
filtered.append(t)
request.tools = filtered if filtered else None
if (request.enable_universal is True) or (
request.enable_universal is None and (request.auto_universal if request.auto_universal is not None else CONFIG.AUTO_ENABLE_TOOLS and detection.get('need_universal'))
):
if not request.tools:
request.tools = [{"name": "universal", "args": {"query": raw_prompt}}]
executed_tool_calls = []
prompt, executed_tool_calls = execute_tools(request, detection, prompt, executed_tool_calls, web_search_enabled, tools_enabled, file_tool_enabled, raw_prompt)
logger.info(f"[REQ] {completionId} - prompt - {prompt}")
# Resume or prefill tokens/state
if request.state_name:
state_key = (request.model, request.state_name)
if state_key in STATE_STORE:
stored = STATE_STORE[state_key]
model_state = stored.get('state', None)
model_tokens = stored.get('model_tokens', [0])
if model_state is None:
# Recompute out and model_state from tokens since we did not persist the torch state
out, model_state = _recompute_out_and_state_from_tokens(request.model, model_tokens)
else:
# If we have a model_state, we still need out logits. Compute from last window of tokens
out, _ = _recompute_out_and_state_from_tokens(request.model, model_tokens[-CONFIG.CHUNK_LEN :])
else:
out, model_tokens, model_state = await runPrefill(request, prompt, [0], model_state)
else:
out, model_tokens, model_state = await runPrefill(request, prompt, [0], model_state)
prefillTime = time.time()
promptTokenCount = len(model_tokens)
fullResponse = " <think" if enableReasoning else ""
completionTokenCount = 0
finishReason = None
# Limit model-initiated tool calls per request to avoid loops
model_initiated_tool_calls = 0
MODEL_MAX_TOOL_CALLS = 3
should_restart = True
while should_restart:
should_restart = False
# Compute a bounded max_tokens for generation based on request and global limits
max_gen_tokens = (
getattr(CONFIG, 'MAX_GENERATION_TOKENS_LIMIT', 64000)
if "max_tokens" not in request.model_fields_set and enableReasoning
else (request.max_tokens or 2048)
)
max_tokens_limit = getattr(CONFIG, 'MAX_TOKENS_PER_REQUEST', None)
if max_tokens_limit:
max_gen_tokens = min(max_gen_tokens, max_tokens_limit)
gen = generate(request, out, model_tokens, model_state, max_tokens=max_gen_tokens)
for chunk in gen:
# chunk['content'] is now expected to be a single token's decoded text
fullResponse += chunk["content"]
# Detect model-issued tool call markers within the output
if model_initiated_tool_calls < MODEL_MAX_TOOL_CALLS:
m = TOOL_CALL_RE.search(fullResponse)
if m:
try:
payload_raw = m.group(1)
import json
payload = json.loads(payload_raw)
tool_name = payload.get('name')
tool_args = payload.get('args', {})
tool_res = None
if tool_name == 'web_search':
from utils import web_search
q = tool_args.get('query') or (request.prompt if request.prompt else cleanMessages(request.messages or []))
k = int(tool_args.get('top_k') or request.search_top_k or 3)
tool_res = web_search(q, k)
elif tool_name in ('calc', 'calculator'):
from utils import calc
expr = tool_args.get('expression')
if expr:
tool_res = calc(expr)
else:
try:
tool_res = universal_tool({'query': tool_args.get('query') or payload.get('query') or ''}, allow_web_search=bool(web_search_enabled), allow_tools=bool(tools_enabled), allow_file_tool=bool(file_tool_enabled))
except Exception:
tool_res = None
if tool_res:
if not isinstance(tool_res, dict):
if tool_name in ('calc', 'calculator'):
tool_res_struct = {"action": "calc", "result": str(tool_res), "metadata": {"expression": tool_args.get('expression'), "confidence": 0.98}}
elif tool_name == 'web_search':
tool_res_struct = {"action": "web_search", "result": str(tool_res), "metadata": {"query": tool_args.get('query'), "top_k": tool_args.get('top_k') or request.search_top_k or 3, "confidence": 0.9}}
else:
tool_res_struct = {"action": tool_name, "result": str(tool_res), "metadata": {"confidence": 0.6}}
else:
tool_res_struct = tool_res
exec_entry = {"name": tool_name, "args": tool_args, "result": tool_res_struct, 'initiated_by_model': True}
executed_tool_calls.append(exec_entry)
delta_text = f"ToolResults:\n{tool_res_struct.get('result')}\n\n"
prompt = delta_text + prompt
fullResponse = TOOL_CALL_RE.sub('', fullResponse)
buffer = [fullResponse]
out, model_tokens, model_state = await runPrefill(request, delta_text, model_tokens, model_state)
model_initiated_tool_calls += 1
except Exception as e:
logger.info(f"Model-initiated tool handling error: {e}")
# Check stop sequences (multi-token) after each token
for stop_words in request.stop or []:
if stop_words in fullResponse:
finishReason = f"stop:words:{stop_words}"
break
completionTokenCount += 1
if chunk["finish_reason"]:
finishReason = chunk["finish_reason"]
generateTime = time.time()
responseLog = {
"content": fullResponse,
"finish": finishReason,
"prefill_len": promptTokenCount,
"prefill_tps": round(promptTokenCount / (prefillTime - createTimestamp), 2),
"gen_len": completionTokenCount,
"gen_tps": round(completionTokenCount / (generateTime - prefillTime) if generateTime!=prefillTime else 0, 2),
}
logger.info(f"[RES] {completionId} - {responseLog}")
reasoning_content, content = parse_think_response(fullResponse)
# Apply bias mitigation to the final assistant content
try:
from utils import bias_mitigation
mitigation = bias_mitigation(content)
if mitigation and isinstance(mitigation, dict):
if mitigation.get('suppressed'):
executed_tool_calls.append({"name": "safety_mitigation", "args": {}, "result": {"action": "safety", "result": mitigation.get('sanitized'), "metadata": {"reason": mitigation.get('reason')}}})
content = mitigation.get('sanitized')
except Exception:
pass
response = ChatCompletion(
id=completionId,
created=int(createTimestamp),
model=request.model,
usage=Usage(
prompt_tokens=promptTokenCount,
completion_tokens=completionTokenCount,
total_tokens=promptTokenCount + completionTokenCount,
prompt_tokens_details=PromptTokensDetails(cached_tokens=0),
),
choices=[
ChatCompletionChoice(
index=0,
message=ChatCompletionMessage(
role="Assistant",
content=content,
reasoning_content=reasoning_content if reasoning_content else None,
tool_calls=executed_tool_calls if executed_tool_calls else None,
),
logprobs=None,
finish_reason=finishReason,
)
],
)
# Save state if requested for future resumption
try:
if request.state_name:
STATE_STORE[(request.model, request.state_name)] = {
'state': model_state,
'model_tokens': model_tokens,
}
if getattr(CONFIG, 'STATE_STORE_SAVE_ON_UPDATE', False):
try:
_save_state_store_to_disk(force=True)
except Exception:
pass
except Exception:
pass
return response
async def chatResponseStream(
request: ChatCompletionRequest,
model_state: Any,
completionId: str,
enableReasoning: bool,
):
createTimestamp = int(time.time())
raw_prompt = request.prompt.strip() if request.prompt is not None else cleanMessages(request.messages or [], False)
# Intent detection and defaults: check whether to auto-enable tools, web_search, reasoning
detection = detect_tools_and_reasoning(raw_prompt)
# Resolve flags via central helper (request defaults to None -> auto-detect)
flags = resolve_request_flags(request, detection)
web_search_enabled = flags['web_search_enabled']
tools_enabled = flags['tools_enabled']
file_tool_enabled = flags['file_tool_enabled']
reasoning_enabled = flags['reasoning_enabled']
enableReasoning = reasoning_enabled
try:
ms_cfg = MODEL_STORAGE.get(request.model)
if ms_cfg and ms_cfg.MODEL_CONFIG and hasattr(ms_cfg.MODEL_CONFIG, 'ALLOW_REASONING') and not ms_cfg.MODEL_CONFIG.ALLOW_REASONING:
enableReasoning = False
except Exception:
pass
# file_tool_enabled is derived from resolve_request_flags as well
# Build final prompt after deciding enableReasoning
prompt = raw_prompt if request.prompt is not None else f"{cleanMessages(request.messages or [], enableReasoning)}\n\nAssistant:{' <think' if enableReasoning else ''}"
if tools_enabled and not request.tools:
if detection.get('detected_tools'):
# Filter detected tools by model ALLOW_* flags
detected = detection.get('detected_tools') or []
filtered = []
for t in detected:
tname = t.get('name')
args = t.get('args', {})
ms_check = MODEL_STORAGE.get(request.model)
allowed = True
try:
if ms_check and ms_check.MODEL_CONFIG:
if tname in ('web_search', 'fetch_url') and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_FETCH_URL') and not ms_check.MODEL_CONFIG.ALLOW_FETCH_URL:
allowed = False
if tname == 'summarize' and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_SUMMARIZE') and not ms_check.MODEL_CONFIG.ALLOW_SUMMARIZE:
allowed = False
if tname in ('keywords',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_KEYWORDS') and not ms_check.MODEL_CONFIG.ALLOW_KEYWORDS:
allowed = False
if tname in ('sentiment',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_SENTIMENT') and not ms_check.MODEL_CONFIG.ALLOW_SENTIMENT:
allowed = False
if tname in ('translate',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_TRANSLATE') and not ms_check.MODEL_CONFIG.ALLOW_TRANSLATE:
allowed = False
if tname in ('spell_check',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_SPELL_CHECK') and not ms_check.MODEL_CONFIG.ALLOW_SPELL_CHECK:
allowed = False
if tname in ('format_code',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_FORMAT_CODE') and not ms_check.MODEL_CONFIG.ALLOW_FORMAT_CODE:
allowed = False
if tname in ('explain_code',) and hasattr(ms_check.MODEL_CONFIG, 'ALLOW_EXPLAIN_CODE') and not ms_check.MODEL_CONFIG.ALLOW_EXPLAIN_CODE:
allowed = False
except Exception:
pass
if allowed:
filtered.append(t)
request.tools = filtered if filtered else None
executed_tool_calls = []
if request.tools:
try:
for tool in request.tools:
name = tool.get('name')
args = tool.get('args', {})
if name == 'web_search':
from utils import web_search
search_q = args.get('query') or (request.prompt if request.prompt else cleanMessages(request.messages or []))
search_top_k = int(args.get('top_k') or request.search_top_k or 3)
search_str = web_search(search_q, search_top_k)
if search_str:
search_res_struct = {"action": "web_search", "result": str(search_str), "metadata": {"query": search_q, "top_k": search_top_k, "confidence": 0.9}}
executed_tool_calls.append({"name": "web_search", "args": {"query": search_q, "top_k": search_top_k}, "result": search_res_struct})
prompt = (f"WebSearchResults:\n{search_res_struct.get('result')}\n\n" + prompt)
elif name == 'calc' or name == 'calculator':
from utils import calc
expr = args.get('expression')
if expr:
calc_res = calc(expr)
calc_res_struct = {"action": "calc", "result": str(calc_res), "metadata": {"expression": expr, "confidence": 0.98}}
executed_tool_calls.append({"name": "calc", "args": {"expression": expr}, "result": calc_res_struct})
prompt = (f"CalcResult:{expr} = {calc_res_struct.get('result')}\n\n" + prompt)
elif name == 'universal':
try:
res = universal_tool(args or {"query": raw_prompt}, allow_web_search=bool(web_search_enabled), allow_tools=bool(tools_enabled), allow_file_tool=bool(file_tool_enabled))
if isinstance(res, dict):
result_text = res.get('result') if res.get('result') is not None else ''
else:
result_text = str(res)
executed_tool_calls.append({"name": "universal", "args": args, "result": res})
prompt = (f"ToolResults:\n{result_text}\n\n" + prompt)
except Exception as e:
logger.info(f"Universal tool execution error: {e}")
else:
logger.info(f"Unsupported tool requested: {name}")
if name == 'fetch_url' or name == 'get_url':
try:
from utils import fetch_url
url = args.get('url') or args.get('uri') or (request.prompt if request.prompt else cleanMessages(request.messages or []))
if not url:
continue
page = fetch_url(url, int(args.get('max_chars') or 20000))
executed_tool_calls.append({"name": "fetch_url", "args": {"url": url}, "result": {"action": "fetch_url", "result": page, "metadata": {"url": url}}})
prompt = (f"ToolResults:\n{page}\n\n" + prompt)
except Exception as e:
logger.info(f"fetch_url tool error: {e}")
if name == 'summarize' or name == 'summary':
try:
from utils import summarize_text, fetch_url
txt = args.get('text') or ''
if not txt and args.get('url'):
txt = fetch_url(args.get('url'))
if not txt and request.prompt:
txt = request.prompt
if not txt and request.messages:
txt = cleanMessages(request.messages or [])
if txt:
s = summarize_text(txt, int(args.get('max_sentences') or 3))
executed_tool_calls.append({"name": "summarize", "args": args, "result": {"action": "summarize", "result": s, "metadata": {"confidence": 0.85}}})
prompt = (f"ToolResults:\n{s}\n\n" + prompt)
except Exception as e:
logger.info(f"summarize tool error: {e}")
if name == 'keywords' or name == 'keyword_extraction':
try:
from utils import extract_keywords, fetch_url
txt = args.get('text') or ''
if not txt and args.get('url'):
txt = fetch_url(args.get('url'))
if not txt and request.prompt:
txt = request.prompt
kws = extract_keywords(txt, int(args.get('top_k') or 5))
executed_tool_calls.append({"name": "keywords", "args": args, "result": {"action": "keywords", "result": kws, "metadata": {"top_k": int(args.get('top_k') or 5), "confidence": 0.78}}})
prompt = (f"ToolResults:\nKeywords:{','.join(kws)}\n\n" + prompt)
except Exception as e:
logger.info(f"keywords tool error: {e}")
if name == 'sentiment' or name == 'tone':
try:
from utils import sentiment_analysis, fetch_url
txt = args.get('text') or ''
if not txt and args.get('url'):
txt = fetch_url(args.get('url'))
if not txt and request.prompt:
txt = request.prompt
res = sentiment_analysis(txt)
executed_tool_calls.append({"name": "sentiment", "args": args, "result": {"action": "sentiment", "result": res, "metadata": {"confidence": res.get('score', 0)}}})
prompt = (f"ToolResults:\nSentiment: {res.get('sentiment')} (score={res.get('score')})\n\n" + prompt)
except Exception as e:
logger.info(f"sentiment tool error: {e}")
except Exception as e:
logger.info(f"Tool processing error: {e}")
elif request.web_search or web_search_enabled:
try:
from utils import web_search
search_q = request.prompt if request.prompt else cleanMessages(request.messages or [])
search_res = web_search(search_q, int(request.search_top_k or 3))
if search_res:
search_res_struct = {"action": "web_search", "result": str(search_res), "metadata": {"query": search_q, "top_k": int(request.search_top_k or 3), "confidence": 0.9}}
executed_tool_calls.append({"name": "web_search", "args": {"query": search_q, "top_k": int(request.search_top_k or 3)}, "result": search_res_struct})
prompt = f"WebSearchResults:\n{search_res_struct.get('result')}\n\n" + prompt
except Exception:
pass
logger.info(f"[REQ] {completionId} - context\n```{prompt}```")
# Resume or prefill tokens/state
if request.state_name:
state_key = (request.model, request.state_name)
if state_key in STATE_STORE:
stored = STATE_STORE[state_key]
model_state = stored.get('state', None)
model_tokens = stored.get('model_tokens', [0])
if model_state is None:
# Recompute out and model_state from tokens since we did not persist the torch state
out, model_state = _recompute_out_and_state_from_tokens(request.model, model_tokens)
else:
out, _ = _recompute_out_and_state_from_tokens(request.model, model_tokens[-CONFIG.CHUNK_LEN :])
else:
out, model_tokens, model_state = await runPrefill(request, prompt, [0], model_state)
else:
out, model_tokens, model_state = await runPrefill(request, prompt, [0], model_state)
prefillTime = time.time()
promptTokenCount = len(model_tokens)
completionTokenCount = 0
finishReason = None
# Limit how many tool calls the model can initiate during a single stream
model_initiated_tool_calls = 0
MODEL_MAX_TOOL_CALLS = 3
response = ChatCompletionChunk(
id=completionId,
created=createTimestamp,
model=request.model,
usage=(
Usage(
prompt_tokens=promptTokenCount,
completion_tokens=completionTokenCount,
total_tokens=promptTokenCount + completionTokenCount,
prompt_tokens_details=PromptTokensDetails(cached_tokens=0),
)
if request.include_usage
else None
),
choices=[
ChatCompletionChoice(
index=0,
delta=ChatCompletionMessage(
role="Assistant",
content="",
reasoning_content="" if enableReasoning else None,
tool_calls=None,
),
logprobs=None,
finish_reason=finishReason,
)
],
)
if response.choices and response.choices[0].delta is None:
response.choices[0].delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
# Attach state_name in the initial chunk so client can save it to continue later
r_dict = response.model_dump()
r_dict['state_name'] = request.state_name
# Attach executed tool_calls both at root for easy client metadata, and within the assistant message delta
if executed_tool_calls:
r_dict['tool_calls'] = executed_tool_calls
try:
if r_dict.get('choices') and len(r_dict['choices']) > 0 and r_dict['choices'][0].get('delta') is not None:
r_dict['choices'][0]['delta']['tool_calls'] = executed_tool_calls
except Exception:
pass
yield f"data: {r_dict}\n\n"
buffer = []
if enableReasoning:
buffer.append("<think")
streamConfig = {
"isChecking": False, # check whether is <think> tag
"fullTextCursor": 0,
"in_think": False,
"cacheStr": "",
}
max_gen_tokens = (
getattr(CONFIG, 'MAX_GENERATION_TOKENS_LIMIT', 64000)
if "max_tokens" not in request.model_fields_set and enableReasoning
else (request.max_tokens or 2048)
)
max_tokens_limit = getattr(CONFIG, 'MAX_TOKENS_PER_REQUEST', None)
if max_tokens_limit:
max_gen_tokens = min(max_gen_tokens, max_tokens_limit)
for chunk in generate(request, out, model_tokens, model_state, max_tokens=max_gen_tokens):
completionTokenCount += 1
# Each token stream is delivered as a decoded character/bytes (maybe 1 or more chars)
chunkContent: str = chunk["content"]
buffer.append(chunkContent)
fullText = "".join(buffer)
if chunk["finish_reason"]:
finishReason = chunk["finish_reason"]
response = ChatCompletionChunk(
id=completionId,
created=createTimestamp,
model=request.model,
usage=(
Usage(
prompt_tokens=promptTokenCount,
completion_tokens=completionTokenCount,
total_tokens=promptTokenCount + completionTokenCount,
prompt_tokens_details=PromptTokensDetails(cached_tokens=0),
)
if request.include_usage
else None
),
choices=[
ChatCompletionChoice(
index=0,
delta=ChatCompletionMessage(
role="Assistant",
content=None,
reasoning_content=None,
tool_calls=None,
),
logprobs=None,
finish_reason=finishReason,
)
],
)
if response.choices and response.choices[0].delta is None:
response.choices[0].delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
markStart = fullText.find("<", streamConfig["fullTextCursor"])
if not streamConfig["isChecking"] and markStart != -1:
streamConfig["isChecking"] = True
if streamConfig["in_think"]:
delta = response.choices[0].delta
if delta is None:
delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
response.choices[0].delta = delta
delta.reasoning_content = fullText[streamConfig["fullTextCursor"] : markStart]
else:
delta = response.choices[0].delta
if delta is None:
delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
response.choices[0].delta = delta
delta.content = fullText[streamConfig["fullTextCursor"] : markStart]
streamConfig["cacheStr"] = ""
streamConfig["fullTextCursor"] = markStart
if streamConfig["isChecking"]:
streamConfig["cacheStr"] = fullText[streamConfig["fullTextCursor"] :]
else:
if streamConfig["in_think"]:
delta = response.choices[0].delta
if delta is None:
delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
response.choices[0].delta = delta
delta.reasoning_content = chunkContent
else:
delta = response.choices[0].delta
if delta is None:
delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
response.choices[0].delta = delta
delta.content = chunkContent
streamConfig["fullTextCursor"] = len(fullText)
markEnd = fullText.find(">", streamConfig["fullTextCursor"])
if (streamConfig["isChecking"] and markEnd != -1) or finishReason != None:
streamConfig["isChecking"] = False
if (
not streamConfig["in_think"]
and streamConfig["cacheStr"].find("<think>") != -1
):
streamConfig["in_think"] = True
delta = response.choices[0].delta
if delta is None:
delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
response.choices[0].delta = delta
delta.reasoning_content = (
delta.reasoning_content
if delta.reasoning_content != None
else "" + streamConfig["cacheStr"].replace("<think>", "")
)
elif (
streamConfig["in_think"]
and streamConfig["cacheStr"].find("</think>") != -1
):
streamConfig["in_think"] = False
delta = response.choices[0].delta
if delta is None:
delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
response.choices[0].delta = delta
delta.content = (
delta.content
if delta.content != None
else "" + streamConfig["cacheStr"].replace("</think>", "")
)
else:
if streamConfig["in_think"]:
delta = response.choices[0].delta
if delta is None:
delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
response.choices[0].delta = delta
delta.reasoning_content = (
delta.reasoning_content
if delta.reasoning_content != None
else "" + streamConfig["cacheStr"]
)
else:
delta = response.choices[0].delta
if delta is None:
delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
response.choices[0].delta = delta
delta.content = (
delta.content
if delta.content != None
else "" + streamConfig["cacheStr"]
)
streamConfig["fullTextCursor"] = len(fullText)
delta = response.choices[0].delta
if delta is None:
delta = ChatCompletionMessage(role="Assistant", content="", reasoning_content=None, tool_calls=None)
response.choices[0].delta = delta
if delta.content != None or delta.reasoning_content != None:
# Save model state frequently (after each token) to allow resuming
try:
if request.state_name:
STATE_STORE[(request.model, request.state_name)] = {
'state': model_state,
'model_tokens': model_tokens,
}
if getattr(CONFIG, 'STATE_STORE_SAVE_ON_UPDATE', False):
try:
_save_state_store_to_disk(force=True)
except Exception:
pass
except Exception:
pass
# model-initiated tool call detection
if model_initiated_tool_calls < MODEL_MAX_TOOL_CALLS:
m = TOOL_CALL_RE.search(fullText)
if m:
try:
payload_raw = m.group(1)
import json
payload = json.loads(payload_raw)
tool_name = payload.get('name')
tool_args = payload.get('args', {})
tool_res = None
if tool_name == 'web_search':
from utils import web_search
q = tool_args.get('query') or (request.prompt if request.prompt else cleanMessages(request.messages or []))
k = int(tool_args.get('top_k') or request.search_top_k or 3)
tool_res = web_search(q, k)
elif tool_name in ('calc', 'calculator'):
from utils import calc
expr = tool_args.get('expression')
if expr:
tool_res = calc(expr)
else:
try:
tool_res = universal_tool({'query': tool_args.get('query') or payload.get('query') or ''}, allow_web_search=bool(web_search_enabled), allow_tools=bool(tools_enabled), allow_file_tool=bool(file_tool_enabled))
except Exception:
tool_res = None
if tool_res:
# Normalize tool_res into a structured dict if needed
if not isinstance(tool_res, dict):
if tool_name in ('calc', 'calculator'):
tool_res_struct = {"action": "calc", "result": str(tool_res), "metadata": {"expression": tool_args.get('expression'), "confidence": 0.98}}
elif tool_name == 'web_search':
tool_res_struct = {"action": "web_search", "result": str(tool_res), "metadata": {"query": tool_args.get('query'), "top_k": tool_args.get('top_k') or request.search_top_k or 3, "confidence": 0.9}}
else:
tool_res_struct = {"action": tool_name, "result": str(tool_res), "metadata": {"confidence": 0.6}}
else:
tool_res_struct = tool_res
exec_entry = {"name": tool_name, "args": tool_args, "result": tool_res_struct, 'initiated_by_model': True}
executed_tool_calls.append(exec_entry)
delta_text = f"ToolResults:\n{tool_res_struct.get('result')}\n\n"
prompt = delta_text + prompt
fullText = TOOL_CALL_RE.sub('', fullText)
buffer = [fullText]
out, model_tokens, model_state = await runPrefill(request, delta_text, model_tokens, model_state)
model_initiated_tool_calls += 1
should_restart = True
break
except Exception as e:
logger.info(f"Model-initiated tool handling error: {e}")
yield f"data: {response.model_dump_json()}\n\n"
# check stop sequences and stop streaming if we see them
for stop_words in request.stop or []:
if stop_words in ''.join(buffer):
finishReason = f"stop:words:{stop_words}"
return
await asyncio.sleep(0)
del streamConfig
else:
should_restart = True
while should_restart:
should_restart = False
gen = generate(request, out, model_tokens, model_state)
for chunk in gen:
completionTokenCount += 1
buffer.append(chunk["content"])
if chunk["finish_reason"]:
finishReason = chunk["finish_reason"]
# Save model state frequently (after each token) to allow resuming
try:
if request.state_name:
STATE_STORE[(request.model, request.state_name)] = {
'state': model_state,
'model_tokens': model_tokens,
}
if getattr(CONFIG, 'STATE_STORE_SAVE_ON_UPDATE', False):
try:
_save_state_store_to_disk(force=True)
except Exception:
pass
except Exception:
pass
# Detect model-initiated tool calls
if model_initiated_tool_calls < MODEL_MAX_TOOL_CALLS:
fullText = ''.join(buffer)
m = TOOL_CALL_RE.search(fullText)
if m:
try:
payload_raw = m.group(1)
import json
payload = json.loads(payload_raw)
tool_name = payload.get('name')
tool_args = payload.get('args', {})
tool_res = None
if tool_name == 'web_search':
from utils import web_search
q = tool_args.get('query') or (request.prompt if request.prompt else cleanMessages(request.messages or []))
k = int(tool_args.get('top_k') or request.search_top_k or 3)
tool_res = web_search(q, k)
elif tool_name in ('calc', 'calculator'):
from utils import calc
expr = tool_args.get('expression')
if expr:
tool_res = calc(expr)
else:
try:
tool_res = universal_tool({'query': tool_args.get('query') or payload.get('query') or ''}, allow_web_search=bool(web_search_enabled), allow_tools=bool(tools_enabled), allow_file_tool=bool(file_tool_enabled))
except Exception:
tool_res = None
if tool_res:
if not isinstance(tool_res, dict):
if tool_name in ('calc', 'calculator'):
tool_res_struct = {"action": "calc", "result": str(tool_res), "metadata": {"expression": tool_args.get('expression'), "confidence": 0.98}}
elif tool_name == 'web_search':
tool_res_struct = {"action": "web_search", "result": str(tool_res), "metadata": {"query": tool_args.get('query'), "top_k": tool_args.get('top_k') or request.search_top_k or 3, "confidence": 0.9}}
else:
tool_res_struct = {"action": tool_name, "result": str(tool_res), "metadata": {"confidence": 0.6}}
else:
tool_res_struct = tool_res
exec_entry = {"name": tool_name, "args": tool_args, "result": tool_res_struct, 'initiated_by_model': True}
executed_tool_calls.append(exec_entry)
delta_text = f"ToolResults:\n{tool_res_struct.get('result')}\n\n"
prompt = delta_text + prompt
fullText = TOOL_CALL_RE.sub('', fullText)
buffer = [fullText]
out, model_tokens, model_state = await runPrefill(request, delta_text, model_tokens, model_state)
# Notify client that a tool was called mid-stream (metadata-only chunk)
try:
meta_resp = ChatCompletionChunk(
id=completionId,
created=createTimestamp,
model=request.model,
usage=(
Usage(
prompt_tokens=promptTokenCount,
completion_tokens=completionTokenCount,
total_tokens=promptTokenCount + completionTokenCount,
prompt_tokens_details=PromptTokensDetails(cached_tokens=0),
)
if request.include_usage
else None
),
choices=[
ChatCompletionChoice(
index=0,
delta=ChatCompletionMessage(role="Assistant", content=None, reasoning_content=None, tool_calls=executed_tool_calls),
logprobs=None,
finish_reason=None,
)
],
)
yield f"data: {meta_resp.model_dump_json()}\n\n"
except Exception:
pass
model_initiated_tool_calls += 1
should_restart = True
break
except Exception as e:
logger.info(f"Model-initiated tool handling error: {e}")
response = ChatCompletionChunk(
id=completionId,
created=createTimestamp,
model=request.model,
usage=(
Usage(
prompt_tokens=promptTokenCount,
completion_tokens=completionTokenCount,
total_tokens=promptTokenCount + completionTokenCount,
prompt_tokens_details=PromptTokensDetails(cached_tokens=0),
)
if request.include_usage
else None
),
choices=[
ChatCompletionChoice(
index=0,
delta=ChatCompletionMessage(role="Assistant", content=chunk["content"], reasoning_content=None, tool_calls=None),
logprobs=None,
finish_reason=finishReason,
)
],
)
yield f"data: {response.model_dump_json()}\n\n"
await asyncio.sleep(0)
genenrateTime = time.time()
responseLog = {
"content": "".join(buffer),
"finish": finishReason,
"prefill_len": promptTokenCount,
"prefill_tps": round(promptTokenCount / (prefillTime - createTimestamp), 2),
"gen_len": completionTokenCount,
"gen_tps": round(completionTokenCount / (genenrateTime - prefillTime), 2),
}
logger.info(f"[RES] {completionId} - {responseLog}")
# Apply bias mitigation to the final content for streaming mode
try:
from utils import bias_mitigation
content_for_mitigation = responseLog.get('content')
if content_for_mitigation is None:
content_for_mitigation = ""
mitigation = bias_mitigation(content_for_mitigation)
if mitigation and isinstance(mitigation, dict) and mitigation.get('suppressed'):
executed_tool_calls.append({"name": "safety_mitigation", "args": {}, "result": {"action": "safety", "result": mitigation.get('sanitized'), "metadata": {"reason": mitigation.get('reason')}}})
responseLog['content'] = mitigation.get('sanitized')
except Exception:
pass
if request.messages is None:
request.messages = []
# Ensure responseLog['content'] is a string
content_str = responseLog["content"] if responseLog["content"] is not None else ""
request.messages.append(ChatMessage(role="Assistant", content=content_str))
log(
{
**request.model_dump(),
**responseLog,
"completionId": completionId,
"machineLabel": os.environ.get("MACHINE_LABEL"),
}
)
del buffer
yield "data: [DONE]\n\n"
@app.post("/api/v1/chat/completions")
async def chat_completions(request: ChatCompletionRequest):
completionId = str(next(CompletionIdGenerator))
logger.info(f"[REQ] {completionId} - {request.model_dump()}")
# Apply any legacy model suffix tags (e.g., 'rwkv-latest:thinking' -> enable_reasoning)
# This helper is defined at module level so it can be unit-tested and reused.
# Apply legacy tags (if present) to request and proceed normally
apply_model_tags_to_request(request)
modelName = request.model
if request.model == "rwkv-latest":
# Map to the default chat model in all cases. Do not redirect to a separate
# reasoning model; the same model will be used and reasoning is handled in-process.
if DEFALUT_MODEL_NAME == None:
raise HTTPException(404, "DEFALUT_MODEL_NAME not set")
ms_def = MODEL_STORAGE.get(DEFALUT_MODEL_NAME)
if not ms_def or not ms_def.MODEL_CONFIG:
raise HTTPException(500, "Default sampler config missing for default model")
defaultSamplerConfig = ms_def.MODEL_CONFIG.DEFAULT_SAMPLER
request.model = DEFALUT_MODEL_NAME
elif modelName in MODEL_STORAGE:
ms_sel = MODEL_STORAGE.get(modelName)
if not ms_sel or not ms_sel.MODEL_CONFIG:
raise HTTPException(500, f"Default sampler config missing for model {modelName}")
defaultSamplerConfig = ms_sel.MODEL_CONFIG.DEFAULT_SAMPLER
request.model = modelName
else:
raise HTTPException(404, f"Can not find `{modelName}`")
# Baseline enableReasoning: prefer explicit request flag if present, else False. chatResponse will recompute with auto-detection.
enableReasoning = bool(request.enable_reasoning) if request.enable_reasoning is not None else False
async def chatResponseStreamDisconnect():
logGPUState()
# Load or initialize model_state and tokens based on state_name
model_state = None
model_tokens_for_resume = [0]
state_name = request.state_name
if state_name is None:
state_name = str(uuid.uuid4())
request.state_name = state_name
state_key = (request.model, state_name)
if state_key in STATE_STORE:
stored = STATE_STORE[state_key]
model_state = stored.get('state', None)
model_tokens_for_resume = stored.get('model_tokens', [0])
request_dict = request.model_dump()
# Apply defaults from model's DEFAULT_SAMPLER, optionally overridden by the
# per-request `sampler` object (or legacy sampler_allow_* booleans).
sampler_overrides = request_dict.get('sampler') or {}
for k, v in defaultSamplerConfig.model_dump().items():
# If the request provided a sampler override for this field, use it
if sampler_overrides and k in sampler_overrides and sampler_overrides.get(k) is not None:
request_dict[k] = sampler_overrides.get(k)
continue
if k in request_dict and request_dict[k] is None:
request_dict[k] = v
realRequest = ChatCompletionRequest(**request_dict)
# Ensure stream defaults to configuration value when not explicitly provided
if realRequest.stream is None:
realRequest.stream = CONFIG.DEFAULT_STREAM
# Enforce top-level numeric limits on the realRequest values
try:
max_tokens_limit = getattr(CONFIG, 'MAX_TOKENS_PER_REQUEST', None)
if realRequest.max_tokens is not None and max_tokens_limit is not None:
realRequest.max_tokens = min(realRequest.max_tokens, max_tokens_limit)
if realRequest.temperature is not None:
realRequest.temperature = min(max(realRequest.temperature, getattr(CONFIG, 'MIN_TEMPERATURE', 0.0)), getattr(CONFIG, 'MAX_TEMPERATURE', 2.0))
if realRequest.top_p is not None:
realRequest.top_p = min(max(realRequest.top_p, getattr(CONFIG, 'MIN_TOP_P', 0.0)), getattr(CONFIG, 'MAX_TOP_P', 1.0))
except Exception:
pass
logger.info(f"[REQ] {completionId} - Real - {request.model_dump()}")
if realRequest.stream:
r = StreamingResponse(
chatResponseStream(realRequest, model_state, completionId, enableReasoning),
media_type="text/event-stream",
background=BackgroundTask(chatResponseStreamDisconnect),
)
else:
r = await chatResponse(realRequest, model_state, completionId, enableReasoning)
# Attach state_name to non-streaming response as additional metadata
try:
import json
if isinstance(r, ChatCompletion):
d = r.model_dump()
d['state_name'] = state_name
return d
except Exception:
pass
return r
# We keep the service API-only by default. If a local `dist-frontend` directory
# exists (a built frontend), mount it at `/` so the app can serve a static UI.
if os.path.isdir("dist-frontend"):
logger.info("Static frontend mount enabled: serving dist-frontend at /")
app.mount("/", StaticFiles(directory="dist-frontend", html=True), name="static")
else:
logger.info("Static frontend mount not enabled; `dist-frontend` directory not found")
@app.get('/api/v1/models')
def list_models():
"""Return model configuration summary for clients/UI.
This endpoint returns configured models, their default sampler values, and
ALLOW_* flags so UI clients can build a controls surface based on server
capabilities (web search, tools, reasoning).
"""
out = []
root_defaults = {
'ALLOW_FILE_TOOL_BY_DEFAULT': getattr(CONFIG, 'ALLOW_FILE_TOOL_BY_DEFAULT', True),
'ENABLE_WEB_SEARCH_BY_DEFAULT': getattr(CONFIG, 'ENABLE_WEB_SEARCH_BY_DEFAULT', True),
'ENABLE_REASONING_BY_DEFAULT': getattr(CONFIG, 'ENABLE_REASONING_BY_DEFAULT', True),
'SHOW_WEB_SEARCH_BUTTON_BY_DEFAULT': getattr(CONFIG, 'SHOW_WEB_SEARCH_BUTTON_BY_DEFAULT', True),
'SHOW_FILE_UPLOAD_BUTTON_BY_DEFAULT': getattr(CONFIG, 'SHOW_FILE_UPLOAD_BUTTON_BY_DEFAULT', True),
'SHOW_REASONING_TOGGLE_BY_DEFAULT': getattr(CONFIG, 'SHOW_REASONING_TOGGLE_BY_DEFAULT', True),
'UPLOAD_URL': '/api/v1/files' if getattr(CONFIG, 'ALLOW_PUBLIC_UPLOADS', False) else None,
'ALLOW_PUBLIC_UPLOADS': getattr(CONFIG, 'ALLOW_PUBLIC_UPLOADS', False),
}
for m in CONFIG.MODELS:
out.append(
{
'SERVICE_NAME': m.SERVICE_NAME,
'DEFAULT_CHAT': m.DEFAULT_CHAT,
'DEFAULT_REASONING': m.DEFAULT_REASONING,
'ALLOW_WEB_SEARCH': getattr(m, 'ALLOW_WEB_SEARCH', True),
'ALLOW_TOOLS': getattr(m, 'ALLOW_TOOLS', True),
'ALLOW_REASONING': getattr(m, 'ALLOW_REASONING', True),
'ALLOW_FILE_TOOL': getattr(m, 'ALLOW_FILE_TOOL', True),
'ALLOW_FETCH_URL': getattr(m, 'ALLOW_FETCH_URL', True),
'ALLOW_SUMMARIZE': getattr(m, 'ALLOW_SUMMARIZE', True),
'ALLOW_KEYWORDS': getattr(m, 'ALLOW_KEYWORDS', True),
'ALLOW_SENTIMENT': getattr(m, 'ALLOW_SENTIMENT', True),
'ALLOW_TRANSLATE': getattr(m, 'ALLOW_TRANSLATE', True),
'ALLOW_SPELL_CHECK': getattr(m, 'ALLOW_SPELL_CHECK', True),
'ALLOW_FORMAT_CODE': getattr(m, 'ALLOW_FORMAT_CODE', True),
'ALLOW_EXPLAIN_CODE': getattr(m, 'ALLOW_EXPLAIN_CODE', True),
'SHOW_WEB_SEARCH_BUTTON': getattr(m, 'SHOW_WEB_SEARCH_BUTTON', True),
'SHOW_FILE_UPLOAD_BUTTON': getattr(m, 'SHOW_FILE_UPLOAD_BUTTON', True),
'SHOW_REASONING_TOGGLE': getattr(m, 'SHOW_REASONING_TOGGLE', True),
'SHOW_FETCH_URL_BUTTON': getattr(m, 'SHOW_FETCH_URL_BUTTON', True),
'SHOW_SUMMARIZE_BUTTON': getattr(m, 'SHOW_SUMMARIZE_BUTTON', True),
'SHOW_KEYWORDS_BUTTON': getattr(m, 'SHOW_KEYWORDS_BUTTON', True),
'SHOW_SENTIMENT_BUTTON': getattr(m, 'SHOW_SENTIMENT_BUTTON', True),
'SHOW_TRANSLATE_BUTTON': getattr(m, 'SHOW_TRANSLATE_BUTTON', True),
'SHOW_SPELL_CHECK_BUTTON': getattr(m, 'SHOW_SPELL_CHECK_BUTTON', True),
'SHOW_FORMAT_CODE_BUTTON': getattr(m, 'SHOW_FORMAT_CODE_BUTTON', True),
'SHOW_EXPLAIN_CODE_BUTTON': getattr(m, 'SHOW_EXPLAIN_CODE_BUTTON', True),
'DEFAULT_SAMPLER': m.DEFAULT_SAMPLER.model_dump() if hasattr(m, 'DEFAULT_SAMPLER') else None,
# Convenience info for clients: upload endpoint and root defaults
'UPLOAD_URL': '/api/v1/files' if getattr(CONFIG, 'ALLOW_PUBLIC_UPLOADS', False) else None,
'UPLOAD_ALLOWED_BY_DEFAULT': getattr(CONFIG, 'ALLOW_FILE_TOOL_BY_DEFAULT', True),
}
)
return {'root_defaults': root_defaults, 'models': out}
def upload_file_internal(bytes_data: bytes, filename: Optional[str] = None, model: Optional[str] = None) -> dict:
"""Internal function to register uploaded file into the service memory.
This is NOT exposed as a public HTTP endpoint by default; it's intended to be used by
the server's `universal_tool` or admin-only utilities. Returns saved metadata or raises Exception.
"""
if not getattr(CONFIG, 'ALLOW_FILE_TOOL_BY_DEFAULT', True):
raise HTTPException(403, 'File uploads are disabled by server configuration')
if model:
if model not in MODEL_STORAGE:
raise HTTPException(404, f"Model {model} not found")
ms = MODEL_STORAGE[model]
if ms and ms.MODEL_CONFIG and not getattr(ms.MODEL_CONFIG, 'ALLOW_FILE_TOOL', True):
raise HTTPException(403, f"Model {model} does not allow file uploads")
from utils import save_bytes_to_upload
meta = save_bytes_to_upload(filename or 'uploaded_file', bytes_data)
if meta.get('error'):
raise Exception(meta.get('error'))
UPLOADED_FILES[meta['file_id']] = meta
return meta
def list_files_internal():
return [UploadedFile(**v).model_dump() for v in UPLOADED_FILES.values()]
def get_file_internal(file_id: str, download: bool = False):
if file_id not in UPLOADED_FILES:
raise HTTPException(404, 'File not found')
meta = UPLOADED_FILES[file_id]
if download:
try:
with open(meta['path'], 'rb') as f:
return StreamingResponse(f, media_type='application/octet-stream')
except Exception as e:
raise HTTPException(500, str(e))
return UploadedFile(**meta)
def delete_file_internal(file_id: str):
if file_id not in UPLOADED_FILES:
raise HTTPException(404, 'File not found')
meta = UPLOADED_FILES.pop(file_id)
try:
if os.path.exists(meta['path']):
os.remove(meta['path'])
except Exception:
pass
return {'success': True}
# If public uploads are allowed, add the public endpoints as wrappers to internal helpers
if getattr(CONFIG, 'ALLOW_PUBLIC_UPLOADS', False):
@app.post('/api/v1/files', response_model=FileUploadResponse)
async def upload_file_public(file: UploadFile = File(...), model: Optional[str] = None):
try:
content = await file.read()
# enforce size limit
max_upload_size = getattr(CONFIG, 'MAX_UPLOAD_SIZE_BYTES', None)
if max_upload_size is not None and len(content) > max_upload_size:
raise HTTPException(413, 'File too large')
fname = file.filename if getattr(file, 'filename', None) else 'uploaded_file'
meta = upload_file_internal(content, filename=fname, model=model)
return FileUploadResponse(success=True, file=UploadedFile(**meta))
except HTTPException:
raise
except Exception as e:
raise HTTPException(500, str(e))
@app.get('/api/v1/files')
def list_files_public():
return list_files_internal()
@app.get('/api/v1/files/{file_id}')
def get_file_public(file_id: str, download: bool = False):
return get_file_internal(file_id, download=download)
@app.delete('/api/v1/files/{file_id}')
def delete_file_public(file_id: str):
return delete_file_internal(file_id)
if __name__ == "__main__":
import uvicorn
host = CONFIG.HOST or "127.0.0.1"
port = CONFIG.PORT or 7860
uvicorn.run(app, host=host, port=port)
|