Skip to content

API Reference

High Level API

High-level Python bindings for llama.cpp.

llama_cpp.Llama

High-level Python wrapper for a llama.cpp model.

Source code in llama_cpp/llama.py
 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
class Llama:
    """High-level Python wrapper for a llama.cpp model."""

    __backend_initialized = False

    def __init__(
        self,
        model_path: str,
        *,
        # Model Params
        n_gpu_layers: int = 0,
        main_gpu: int = 0,
        tensor_split: Optional[List[float]] = None,
        vocab_only: bool = False,
        use_mmap: bool = True,
        use_mlock: bool = False,
        # Context Params
        seed: int = llama_cpp.LLAMA_DEFAULT_SEED,
        n_ctx: int = 512,
        n_batch: int = 512,
        n_threads: Optional[int] = None,
        n_threads_batch: Optional[int] = None,
        rope_scaling_type: Optional[int] = llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED,
        rope_freq_base: float = 0.0,
        rope_freq_scale: float = 0.0,
        yarn_ext_factor: float = -1.0,
        yarn_attn_factor: float = 1.0,
        yarn_beta_fast: float = 32.0,
        yarn_beta_slow: float = 1.0,
        yarn_orig_ctx: int = 0,
        mul_mat_q: bool = True,
        f16_kv: bool = True,
        logits_all: bool = False,
        embedding: bool = False,
        # Sampling Params
        last_n_tokens_size: int = 64,
        # LoRA Params
        lora_base: Optional[str] = None,
        lora_scale: float = 1.0,
        lora_path: Optional[str] = None,
        # Backend Params
        numa: bool = False,
        # Chat Format Params
        chat_format: str = "llama-2",
        chat_handler: Optional[llama_chat_format.LlamaChatCompletionHandler] = None,
        # Misc
        verbose: bool = True,
        # Extra Params
        **kwargs,  # type: ignore
    ):
        """Load a llama.cpp model from `model_path`.

        Examples:
            Basic usage

            >>> import llama_cpp
            >>> model = llama_cpp.Llama(
            ...     model_path="path/to/model",
            ... )
            >>> print(model("The quick brown fox jumps ", stop=["."])["choices"][0]["text"])
            the lazy dog

            Loading a chat model

            >>> import llama_cpp
            >>> model = llama_cpp.Llama(
            ...     model_path="path/to/model",
            ...     chat_format="llama-2",
            ... )
            >>> print(model.create_chat_completion(
            ...     messages=[{
            ...         "role": "user",
            ...         "content": "what is the meaning of life?"
            ...     }]
            ... ))

        Args:
            model_path: Path to the model.
            n_gpu_layers: Number of layers to offload to GPU (-ngl). If -1, all layers are offloaded.
            main_gpu: The GPU that is used for scratch and small tensors.
            tensor_split: How split tensors should be distributed across GPUs. If None, the model is not split.
            vocab_only: Only load the vocabulary no weights.
            use_mmap: Use mmap if possible.
            use_mlock: Force the system to keep the model in RAM.
            seed: RNG seed, -1 for random
            n_ctx: Text context, 0 = from model
            n_batch: Prompt processing maximum batch size
            n_threads: Number of threads to use for generation
            n_threads_batch: Number of threads to use for batch processing
            rope_scaling_type: RoPE scaling type, from `enum llama_rope_scaling_type`. ref: https://github.com/ggerganov/llama.cpp/pull/2054
            rope_freq_base: RoPE base frequency, 0 = from model
            rope_freq_scale: RoPE frequency scaling factor, 0 = from model
            yarn_ext_factor: YaRN extrapolation mix factor, negative = from model
            yarn_attn_factor: YaRN magnitude scaling factor
            yarn_beta_fast: YaRN low correction dim
            yarn_beta_slow: YaRN high correction dim
            yarn_orig_ctx: YaRN original context size
            f16_kv: Use fp16 for KV cache, fp32 otherwise
            logits_all: Return logits for all tokens, not just the last token. Must be True for completion to return logprobs.
            embedding: Embedding mode only.
            last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque.
            lora_base: Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model.
            lora_path: Path to a LoRA file to apply to the model.
            numa: Enable NUMA support. (NOTE: The initial value of this parameter is used for the remainder of the program as this value is set in llama_backend_init)
            chat_format: String specifying the chat format to use when calling create_chat_completion.
            chat_handler: Optional chat handler to use when calling create_chat_completion.
            verbose: Print verbose output to stderr.

        Raises:
            ValueError: If the model path does not exist.

        Returns:
            A Llama instance.
        """
        self.verbose = verbose

        self.numa = numa
        if not Llama.__backend_initialized:
            with suppress_stdout_stderr(disable=self.verbose):
                llama_cpp.llama_backend_init(self.numa)
            Llama.__backend_initialized = True

        self.model_path = model_path

        # Model Params
        self.model_params = llama_cpp.llama_model_default_params()
        self.model_params.n_gpu_layers = (
            0x7FFFFFFF if n_gpu_layers == -1 else n_gpu_layers
        )  # 0x7FFFFFFF is INT32 max, will be auto set to all layers
        self.model_params.main_gpu = main_gpu
        self.tensor_split = tensor_split
        self._p_tensor_split = None
        if self.tensor_split is not None:
            if len(self.tensor_split) > llama_cpp.LLAMA_MAX_DEVICES:
                raise ValueError(
                    f"Attempt to split tensors that exceed maximum supported devices. Current LLAMA_MAX_DEVICES={llama_cpp.LLAMA_MAX_DEVICES}"
                )
            # Type conversion and expand the list to the length of LLAMA_MAX_DEVICES
            FloatArray = ctypes.c_float * llama_cpp.LLAMA_MAX_DEVICES
            self._c_tensor_split = FloatArray(
                *tensor_split  # type: ignore
            )  # keep a reference to the array so it is not gc'd
            self.model_params.tensor_split = self._c_tensor_split
        self.model_params.vocab_only = vocab_only
        self.model_params.use_mmap = use_mmap if lora_path is None else False
        self.model_params.use_mlock = use_mlock

        self.n_batch = min(n_ctx, n_batch)  # ???
        self.n_threads = n_threads or max(multiprocessing.cpu_count() // 2, 1)
        self.n_threads_batch = n_threads_batch or max(
            multiprocessing.cpu_count() // 2, 1
        )
        # Context Params
        self.context_params = llama_cpp.llama_context_default_params()
        self.context_params.seed = seed
        self.context_params.n_ctx = n_ctx
        self.context_params.n_batch = self.n_batch
        self.context_params.n_threads = self.n_threads
        self.context_params.n_threads_batch = self.n_threads_batch
        self.context_params.rope_scaling_type = (
            rope_scaling_type
            if rope_scaling_type is not None
            else llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED
        )
        self.context_params.rope_freq_base = (
            rope_freq_base if rope_freq_base != 0.0 else 0
        )
        self.context_params.rope_freq_scale = (
            rope_freq_scale if rope_freq_scale != 0.0 else 0
        )
        self.context_params.yarn_ext_factor = (
            yarn_ext_factor if yarn_ext_factor != 0.0 else 0
        )
        self.context_params.yarn_attn_factor = (
            yarn_attn_factor if yarn_attn_factor != 0.0 else 0
        )
        self.context_params.yarn_beta_fast = (
            yarn_beta_fast if yarn_beta_fast != 0.0 else 0
        )
        self.context_params.yarn_beta_slow = (
            yarn_beta_slow if yarn_beta_slow != 0.0 else 0
        )
        self.context_params.yarn_orig_ctx = yarn_orig_ctx if yarn_orig_ctx != 0 else 0
        self.context_params.mul_mat_q = mul_mat_q
        self.context_params.f16_kv = f16_kv
        self.context_params.logits_all = logits_all
        self.context_params.embedding = embedding

        # Sampling Params
        self.last_n_tokens_size = last_n_tokens_size

        self.cache: Optional[BaseLlamaCache] = None

        self.lora_base = lora_base
        self.lora_scale = lora_scale
        self.lora_path = lora_path

        if not os.path.exists(model_path):
            raise ValueError(f"Model path does not exist: {model_path}")

        self._model = _LlamaModel(
            path_model=self.model_path, params=self.model_params, verbose=self.verbose
        )

        self._ctx = _LlamaContext(
            model=self._model,
            params=self.context_params,
            verbose=self.verbose,
        )

        self._batch = _LlamaBatch(
            n_tokens=self.n_batch,
            embd=0,
            n_seq_max=self.context_params.n_ctx,
            verbose=self.verbose,
        )

        if self.lora_path:
            if self._model.apply_lora_from_file(
                self.lora_path,
                self.lora_scale,
                self.lora_base,
                self.n_threads,
            ):
                raise RuntimeError(
                    f"Failed to apply LoRA from lora path: {self.lora_path} to base path: {self.lora_base}"
                )

        if self.verbose:
            print(llama_cpp.llama_print_system_info().decode("utf-8"), file=sys.stderr)

        self.chat_format = chat_format
        self.chat_handler = chat_handler

        self._n_vocab = self.n_vocab()
        self._n_ctx = self.n_ctx()

        self._token_nl = self.token_nl()
        self._token_eos = self.token_eos()

        self._candidates = _LlamaTokenDataArray(n_vocab=self._n_vocab)

        self.n_tokens = 0
        self.input_ids: npt.NDArray[np.intc] = np.ndarray((n_ctx,), dtype=np.intc)
        self.scores: npt.NDArray[np.single] = np.ndarray(
            (n_ctx, self._n_vocab), dtype=np.single
        )

    @property
    def ctx(self) -> llama_cpp.llama_context_p:
        assert self._ctx.ctx is not None
        return self._ctx.ctx

    @property
    def model(self) -> llama_cpp.llama_model_p:
        assert self._model.model is not None
        return self._model.model

    @property
    def _input_ids(self) -> npt.NDArray[np.intc]:
        return self.input_ids[: self.n_tokens]

    @property
    def _scores(self) -> npt.NDArray[np.single]:
        return self.scores[: self.n_tokens, :]

    @property
    def eval_tokens(self) -> Deque[int]:
        return deque(self.input_ids[: self.n_tokens].tolist(), maxlen=self._n_ctx)

    @property
    def eval_logits(self) -> Deque[List[float]]:
        return deque(
            self.scores[: self.n_tokens, :].tolist(),
            maxlen=self._n_ctx if self.context_params.logits_all else 1,
        )

    def tokenize(
        self, text: bytes, add_bos: bool = True, special: bool = False
    ) -> List[int]:
        """Tokenize a string.

        Args:
            text: The utf-8 encoded string to tokenize.

        Raises:
            RuntimeError: If the tokenization failed.

        Returns:
            A list of tokens.
        """
        return self._model.tokenize(text, add_bos, special)

    def detokenize(self, tokens: List[int]) -> bytes:
        """Detokenize a list of tokens.

        Args:
            tokens: The list of tokens to detokenize.

        Returns:
            The detokenized string.
        """
        return self._model.detokenize(tokens)

    def set_cache(self, cache: Optional[BaseLlamaCache]):
        """Set the cache.

        Args:
            cache: The cache to set.
        """
        self.cache = cache

    def set_seed(self, seed: int):
        """Set the random seed.

        Args:
            seed: The random seed.
        """
        assert self._ctx.ctx is not None
        llama_cpp.llama_set_rng_seed(self._ctx.ctx, seed)

    def reset(self):
        """Reset the model state."""
        self.n_tokens = 0

    def eval(self, tokens: Sequence[int]):
        """Evaluate a list of tokens.

        Args:
            tokens: The list of tokens to evaluate.
        """
        assert self._ctx.ctx is not None
        assert self._batch.batch is not None
        self._ctx.kv_cache_seq_rm(-1, self.n_tokens, -1)
        for i in range(0, len(tokens), self.n_batch):
            batch = tokens[i : min(len(tokens), i + self.n_batch)]
            n_past = self.n_tokens
            n_tokens = len(batch)
            self._batch.set_batch(
                batch=batch, n_past=n_past, logits_all=self.context_params.logits_all
            )
            self._ctx.decode(self._batch)
            # Save tokens
            self.input_ids[n_past : n_past + n_tokens] = batch
            # Save logits
            rows = n_tokens
            cols = self._n_vocab
            offset = (
                0 if self.context_params.logits_all else n_tokens - 1
            )  # NOTE: Only save the last token logits if logits_all is False
            self.scores[n_past + offset : n_past + n_tokens, :].reshape(-1)[
                :
            ] = self._ctx.get_logits()[offset * cols : rows * cols]
            # Update n_tokens
            self.n_tokens += n_tokens

    def sample(
        self,
        top_k: int = 40,
        top_p: float = 0.95,
        min_p: float = 0.05,
        typical_p: float = 1.0,
        temp: float = 0.80,
        repeat_penalty: float = 1.1,
        frequency_penalty: float = 0.0,
        presence_penalty: float = 0.0,
        tfs_z: float = 1.0,
        mirostat_mode: int = 0,
        mirostat_eta: float = 0.1,
        mirostat_tau: float = 5.0,
        penalize_nl: bool = True,
        logits_processor: Optional[LogitsProcessorList] = None,
        grammar: Optional[LlamaGrammar] = None,
    ):
        """Sample a token from the model.

        Args:
            top_k: The top-k sampling parameter.
            top_p: The top-p sampling parameter.
            temp: The temperature parameter.
            repeat_penalty: The repeat penalty parameter.

        Returns:
            The sampled token.
        """
        assert self._ctx is not None
        assert self.n_tokens > 0
        last_n_tokens_data = [llama_cpp.llama_token(0)] * max(
            0, self.last_n_tokens_size - self.n_tokens
        ) + self._input_ids[-self.last_n_tokens_size :].tolist()
        last_n_tokens_size = len(last_n_tokens_data)
        n_vocab = self._n_vocab
        n_ctx = self._n_ctx
        top_k = n_vocab if top_k <= 0 else top_k
        last_n_tokens_size = n_ctx if last_n_tokens_size < 0 else last_n_tokens_size
        last_n_tokens_data_c = (llama_cpp.llama_token * last_n_tokens_size)(
            *last_n_tokens_data
        )
        logits: npt.NDArray[np.single] = self._scores[-1, :]

        if logits_processor is not None:
            logits[:] = logits_processor(self._input_ids, logits)

        nl_logit = logits[self._token_nl]
        self._candidates.copy_logits(logits)
        self._ctx.sample_repetition_penalties(
            candidates=self._candidates,
            last_tokens_data=last_n_tokens_data_c,
            penalty_last_n=last_n_tokens_size,
            penalty_repeat=repeat_penalty,
            penalty_freq=frequency_penalty,
            penalty_present=presence_penalty,
        )
        if not penalize_nl:
            self._candidates.candidates.data[self._token_nl].logit = llama_cpp.c_float(
                nl_logit
            )

        if grammar is not None:
            self._ctx.sample_grammar(
                candidates=self._candidates,
                grammar=grammar,
            )

        if temp < 0.0:
            self._ctx.sample_softmax(candidates=self._candidates)
            id = self._candidates.candidates.data[0].id
        elif temp == 0.0:
            id = self._ctx.sample_token_greedy(candidates=self._candidates)
        elif mirostat_mode == 1:
            self._ctx.sample_temp(candidates=self._candidates, temp=temp)
            id = self._ctx.sample_token_mirostat(
                candidates=self._candidates,
                tau=mirostat_tau,
                eta=mirostat_eta,
                mu=2.0 * mirostat_tau,
                m=100,
            )
        elif mirostat_mode == 2:
            self._ctx.sample_temp(candidates=self._candidates, temp=temp)
            id = self._ctx.sample_token_mirostat_v2(
                candidates=self._candidates,
                tau=mirostat_tau,
                eta=mirostat_eta,
                mu=2.0 * mirostat_tau,
            )
        else:
            self._ctx.sample_top_k(candidates=self._candidates, k=top_k, min_keep=1)
            self._ctx.sample_tail_free(candidates=self._candidates, z=tfs_z, min_keep=1)
            self._ctx.sample_typical(
                candidates=self._candidates, p=typical_p, min_keep=1
            )
            self._ctx.sample_top_p(candidates=self._candidates, p=top_p, min_keep=1)
            self._ctx.sample_min_p(candidates=self._candidates, p=min_p, min_keep=1)
            self._ctx.sample_temp(candidates=self._candidates, temp=temp)
            id = self._ctx.sample_token(candidates=self._candidates)
        if grammar is not None:
            self._ctx.grammar_accept_token(grammar=grammar, token=id)
        return id

    def generate(
        self,
        tokens: Sequence[int],
        top_k: int = 40,
        top_p: float = 0.95,
        min_p: float = 0.05,
        typical_p: float = 1.0,
        temp: float = 0.80,
        repeat_penalty: float = 1.1,
        reset: bool = True,
        frequency_penalty: float = 0.0,
        presence_penalty: float = 0.0,
        tfs_z: float = 1.0,
        mirostat_mode: int = 0,
        mirostat_tau: float = 5.0,
        mirostat_eta: float = 0.1,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        grammar: Optional[LlamaGrammar] = None,
    ) -> Generator[int, Optional[Sequence[int]], None]:
        """Create a generator of tokens from a prompt.

        Examples:
            >>> llama = Llama("models/ggml-7b.bin")
            >>> tokens = llama.tokenize(b"Hello, world!")
            >>> for token in llama.generate(tokens, top_k=40, top_p=0.95, temp=1.0, repeat_penalty=1.1):
            ...     print(llama.detokenize([token]))

        Args:
            tokens: The prompt tokens.
            top_k: The top-k sampling parameter.
            top_p: The top-p sampling parameter.
            temp: The temperature parameter.
            repeat_penalty: The repeat penalty parameter.
            reset: Whether to reset the model state.

        Yields:
            The generated tokens.
        """
        if reset and self.n_tokens > 0:
            longest_prefix = 0
            for a, b in zip(self._input_ids, tokens[:-1]):
                if a == b:
                    longest_prefix += 1
                else:
                    break
            if longest_prefix > 0:
                if self.verbose:
                    print("Llama.generate: prefix-match hit", file=sys.stderr)
                reset = False
                tokens = tokens[longest_prefix:]
                self.n_tokens = longest_prefix

        if reset:
            self.reset()

        if grammar is not None:
            grammar.reset()

        while True:
            self.eval(tokens)
            token = self.sample(
                top_k=top_k,
                top_p=top_p,
                min_p=min_p,
                typical_p=typical_p,
                temp=temp,
                repeat_penalty=repeat_penalty,
                frequency_penalty=frequency_penalty,
                presence_penalty=presence_penalty,
                tfs_z=tfs_z,
                mirostat_mode=mirostat_mode,
                mirostat_tau=mirostat_tau,
                mirostat_eta=mirostat_eta,
                logits_processor=logits_processor,
                grammar=grammar,
            )
            if stopping_criteria is not None and stopping_criteria(
                self._input_ids, self._scores[-1, :]
            ):
                return
            tokens_or_none = yield token
            tokens = [token]
            if tokens_or_none is not None:
                tokens.extend(tokens_or_none)

    def create_embedding(
        self, input: Union[str, List[str]], model: Optional[str] = None
    ) -> CreateEmbeddingResponse:
        """Embed a string.

        Args:
            input: The utf-8 encoded string to embed.

        Returns:
            An embedding object.
        """
        assert self._ctx.ctx is not None
        assert self._model.model is not None
        model_name: str = model if model is not None else self.model_path

        if self.context_params.embedding == False:
            raise RuntimeError(
                "Llama model must be created with embedding=True to call this method"
            )

        if self.verbose:
            llama_cpp.llama_reset_timings(self._ctx.ctx)

        if isinstance(input, str):
            inputs = [input]
        else:
            inputs = input

        data: List[Embedding] = []
        total_tokens = 0
        for index, input in enumerate(inputs):
            tokens = self.tokenize(input.encode("utf-8"), special=True)
            self.reset()
            self.eval(tokens)
            n_tokens = len(tokens)
            total_tokens += n_tokens
            embedding = llama_cpp.llama_get_embeddings(self._ctx.ctx)[
                : llama_cpp.llama_n_embd(self._model.model)
            ]

            data.append(
                {
                    "object": "embedding",
                    "embedding": embedding,
                    "index": index,
                }
            )
        if self.verbose:
            llama_cpp.llama_print_timings(self._ctx.ctx)

        return {
            "object": "list",
            "data": data,
            "model": model_name,
            "usage": {
                "prompt_tokens": total_tokens,
                "total_tokens": total_tokens,
            },
        }

    def embed(self, input: str) -> List[float]:
        """Embed a string.

        Args:
            input: The utf-8 encoded string to embed.

        Returns:
            A list of embeddings
        """
        return list(map(float, self.create_embedding(input)["data"][0]["embedding"]))

    def _create_completion(
        self,
        prompt: Union[str, List[int]],
        suffix: Optional[str] = None,
        max_tokens: Optional[int] = 16,
        temperature: float = 0.8,
        top_p: float = 0.95,
        min_p: float = 0.05,
        typical_p: float = 1.0,
        logprobs: Optional[int] = None,
        echo: bool = False,
        stop: Optional[Union[str, List[str]]] = [],
        frequency_penalty: float = 0.0,
        presence_penalty: float = 0.0,
        repeat_penalty: float = 1.1,
        top_k: int = 40,
        stream: bool = False,
        seed: Optional[int] = None,
        tfs_z: float = 1.0,
        mirostat_mode: int = 0,
        mirostat_tau: float = 5.0,
        mirostat_eta: float = 0.1,
        model: Optional[str] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        logits_processor: Optional[LogitsProcessorList] = None,
        grammar: Optional[LlamaGrammar] = None,
        logit_bias: Optional[Dict[str, float]] = None,
    ) -> Union[
        Iterator[CreateCompletionResponse], Iterator[CreateCompletionStreamResponse]
    ]:
        assert self._ctx is not None
        assert suffix is None or suffix.__class__ is str

        completion_id: str = f"cmpl-{str(uuid.uuid4())}"
        created: int = int(time.time())
        # If prompt is empty, initialize completion with BOS token to avoid
        # detokenization including a space at the beginning of the completion
        completion_tokens: List[int] = [] if len(prompt) > 0 else [self.token_bos()]
        # Add blank space to start of prompt to match OG llama tokenizer
        prompt_tokens: List[int] = (
            (
                self.tokenize(prompt.encode("utf-8"), special=True)
                if prompt != ""
                else [self.token_bos()]
            )
            if isinstance(prompt, str)
            else prompt
        )
        text: bytes = b""
        returned_tokens: int = 0
        stop = (
            stop if isinstance(stop, list) else [stop] if isinstance(stop, str) else []
        )
        model_name: str = model if model is not None else self.model_path

        # NOTE: This likely doesn't work correctly for the first token in the prompt
        # because of the extra space added to the start of the prompt_tokens
        if logit_bias is not None:
            logit_bias_map = {int(k): float(v) for k, v in logit_bias.items()}

            def logit_bias_processor(
                input_ids: npt.NDArray[np.intc],
                scores: npt.NDArray[np.single],
            ) -> npt.NDArray[np.single]:
                new_scores = np.copy(
                    scores
                )  # Does it make sense to copy the whole array or can we just overwrite the original one?
                for input_id, score in logit_bias_map.items():
                    new_scores[input_id] = score + scores[input_id]
                return new_scores

            _logit_bias_processor = LogitsProcessorList([logit_bias_processor])
            if logits_processor is None:
                logits_processor = _logit_bias_processor
            else:
                logits_processor = logits_processor.extend(_logit_bias_processor)

        if self.verbose:
            self._ctx.reset_timings()

        if len(prompt_tokens) >= self._n_ctx:
            raise ValueError(
                f"Requested tokens ({len(prompt_tokens)}) exceed context window of {llama_cpp.llama_n_ctx(self.ctx)}"
            )

        if max_tokens is None or max_tokens <= 0:
            # Unlimited, depending on n_ctx.
            max_tokens = self._n_ctx - len(prompt_tokens)

        # Truncate max_tokens if requested tokens would exceed the context window
        max_tokens = (
            max_tokens
            if max_tokens + len(prompt_tokens) < self._n_ctx
            else (self._n_ctx - len(prompt_tokens))
        )

        if stop != []:
            stop_sequences = [s.encode("utf-8") for s in stop]
        else:
            stop_sequences = []

        if logprobs is not None and self.context_params.logits_all is False:
            raise ValueError(
                "logprobs is not supported for models created with logits_all=False"
            )

        if self.cache:
            try:
                cache_item = self.cache[prompt_tokens]
                cache_prefix_len = Llama.longest_token_prefix(
                    cache_item.input_ids.tolist(), prompt_tokens
                )
                eval_prefix_len = Llama.longest_token_prefix(
                    self._input_ids.tolist(), prompt_tokens
                )
                if cache_prefix_len > eval_prefix_len:
                    self.load_state(cache_item)
                    if self.verbose:
                        print("Llama._create_completion: cache hit", file=sys.stderr)
            except KeyError:
                if self.verbose:
                    print("Llama._create_completion: cache miss", file=sys.stderr)

        if seed is not None:
            self._ctx.set_rng_seed(seed)

        finish_reason = "length"
        multibyte_fix = 0
        for token in self.generate(
            prompt_tokens,
            top_k=top_k,
            top_p=top_p,
            min_p=min_p,
            typical_p=typical_p,
            temp=temperature,
            tfs_z=tfs_z,
            mirostat_mode=mirostat_mode,
            mirostat_tau=mirostat_tau,
            mirostat_eta=mirostat_eta,
            frequency_penalty=frequency_penalty,
            presence_penalty=presence_penalty,
            repeat_penalty=repeat_penalty,
            stopping_criteria=stopping_criteria,
            logits_processor=logits_processor,
            grammar=grammar,
        ):
            if token == self._token_eos:
                text = self.detokenize(completion_tokens)
                finish_reason = "stop"
                break

            completion_tokens.append(token)

            all_text = self.detokenize(completion_tokens)

            # Contains multi-byte UTF8
            for k, char in enumerate(all_text[-3:]):
                k = 3 - k
                for num, pattern in [(2, 192), (3, 224), (4, 240)]:
                    # Bitwise AND check
                    if num > k and pattern & char == pattern:
                        multibyte_fix = num - k

            # Stop incomplete bytes from passing
            if multibyte_fix > 0:
                multibyte_fix -= 1
                continue

            any_stop = [s for s in stop_sequences if s in all_text]
            if len(any_stop) > 0:
                first_stop = any_stop[0]
                text = all_text[: all_text.index(first_stop)]
                finish_reason = "stop"
                break

            if stream:
                remaining_tokens = completion_tokens[returned_tokens:]
                remaining_text = self.detokenize(remaining_tokens)
                remaining_length = len(remaining_text)

                # We want to avoid yielding any characters from
                # the generated text if they are part of a stop
                # sequence.
                first_stop_position = 0
                for s in stop_sequences:
                    for i in range(min(len(s), remaining_length), 0, -1):
                        if remaining_text.endswith(s[:i]):
                            if i > first_stop_position:
                                first_stop_position = i
                            break

                token_end_position = 0

                if logprobs is not None:
                    # not sure how to handle this branch when dealing
                    # with CJK output, so keep it unchanged
                    for token in remaining_tokens:
                        if token == self.token_bos():
                            continue
                        token_end_position += len(self.detokenize([token]))
                        # Check if stop sequence is in the token
                        if token_end_position > (
                            remaining_length - first_stop_position
                        ):
                            break
                        token_str = self.detokenize([token]).decode(
                            "utf-8", errors="ignore"
                        )
                        text_offset = len(prompt) + len(
                            self.detokenize(completion_tokens[:returned_tokens])
                        )
                        token_offset = len(prompt_tokens) + returned_tokens
                        logits = self._scores[token_offset - 1, :].tolist()
                        current_logprobs = Llama.logits_to_logprobs(logits)
                        sorted_logprobs = list(
                            sorted(
                                zip(current_logprobs, range(len(current_logprobs))),
                                reverse=True,
                            )
                        )
                        top_logprob = {
                            self.detokenize([i]).decode(
                                "utf-8", errors="ignore"
                            ): logprob
                            for logprob, i in sorted_logprobs[:logprobs]
                        }
                        top_logprob.update({token_str: current_logprobs[int(token)]})
                        logprobs_or_none = {
                            "tokens": [
                                self.detokenize([token]).decode(
                                    "utf-8", errors="ignore"
                                )
                            ],
                            "text_offset": [text_offset],
                            "token_logprobs": [current_logprobs[int(token)]],
                            "top_logprobs": [top_logprob],
                        }
                        returned_tokens += 1
                        yield {
                            "id": completion_id,
                            "object": "text_completion",
                            "created": created,
                            "model": model_name,
                            "choices": [
                                {
                                    "text": self.detokenize([token]).decode(
                                        "utf-8", errors="ignore"
                                    ),
                                    "index": 0,
                                    "logprobs": logprobs_or_none,
                                    "finish_reason": None,
                                }
                            ],
                        }
                else:
                    while len(remaining_tokens) > 0:
                        decode_success = False
                        for i in range(1, len(remaining_tokens) + 1):
                            try:
                                bs = self.detokenize(remaining_tokens[:i])
                                ts = bs.decode("utf-8")
                                decode_success = True
                                break
                            except UnicodeError:
                                pass
                        else:
                            break
                        if not decode_success:
                            # all remaining tokens cannot be decoded to a UTF-8 character
                            break
                        token_end_position += len(bs)
                        if token_end_position > (
                            remaining_length - first_stop_position
                        ):
                            break
                        remaining_tokens = remaining_tokens[i:]
                        returned_tokens += i

                        yield {
                            "id": completion_id,
                            "object": "text_completion",
                            "created": created,
                            "model": model_name,
                            "choices": [
                                {
                                    "text": ts,
                                    "index": 0,
                                    "logprobs": None,
                                    "finish_reason": None,
                                }
                            ],
                        }

            if len(completion_tokens) >= max_tokens:
                text = self.detokenize(completion_tokens)
                finish_reason = "length"
                break

        if stopping_criteria is not None and stopping_criteria(
            self._input_ids, self._scores[-1, :]
        ):
            text = self.detokenize(completion_tokens)
            finish_reason = "stop"

        if self.verbose:
            self._ctx.print_timings()

        if stream:
            remaining_tokens = completion_tokens[returned_tokens:]
            all_text = self.detokenize(remaining_tokens)
            any_stop = [s for s in stop_sequences if s in all_text]
            if len(any_stop) > 0:
                end = min(all_text.index(stop) for stop in any_stop)
            else:
                end = len(all_text)

            token_end_position = 0
            for token in remaining_tokens:
                token_end_position += len(self.detokenize([token]))

                logprobs_or_none: Optional[CompletionLogprobs] = None
                if logprobs is not None:
                    if token == self.token_bos():
                        continue
                    token_str = self.detokenize([token]).decode(
                        "utf-8", errors="ignore"
                    )
                    text_offset = len(prompt) + len(
                        self.detokenize(completion_tokens[:returned_tokens])
                    )
                    token_offset = len(prompt_tokens) + returned_tokens - 1
                    logits = self._scores[token_offset, :].tolist()
                    current_logprobs = Llama.logits_to_logprobs(logits)
                    sorted_logprobs = list(
                        sorted(
                            zip(current_logprobs, range(len(current_logprobs))),
                            reverse=True,
                        )
                    )
                    top_logprob = {
                        self.detokenize([i]).decode("utf-8", errors="ignore"): logprob
                        for logprob, i in sorted_logprobs[:logprobs]
                    }
                    top_logprob.update({token_str: current_logprobs[int(token)]})
                    logprobs_or_none = {
                        "tokens": [
                            self.detokenize([token]).decode("utf-8", errors="ignore")
                        ],
                        "text_offset": [text_offset],
                        "token_logprobs": [current_logprobs[int(token)]],
                        "top_logprobs": [top_logprob],
                    }

                if token_end_position >= end:
                    last_text = self.detokenize([token])
                    if token_end_position == end - 1:
                        break
                    returned_tokens += 1
                    yield {
                        "id": completion_id,
                        "object": "text_completion",
                        "created": created,
                        "model": model_name,
                        "choices": [
                            {
                                "text": last_text[
                                    : len(last_text) - (token_end_position - end)
                                ].decode("utf-8", errors="ignore"),
                                "index": 0,
                                "logprobs": logprobs_or_none,
                                "finish_reason": None,
                            }
                        ],
                    }
                    break
                returned_tokens += 1
                yield {
                    "id": completion_id,
                    "object": "text_completion",
                    "created": created,
                    "model": model_name,
                    "choices": [
                        {
                            "text": self.detokenize([token]).decode(
                                "utf-8", errors="ignore"
                            ),
                            "index": 0,
                            "logprobs": logprobs_or_none,
                            "finish_reason": None,
                        }
                    ],
                }
            yield {
                "id": completion_id,
                "object": "text_completion",
                "created": created,
                "model": model_name,
                "choices": [
                    {
                        "text": "",
                        "index": 0,
                        "logprobs": None,
                        "finish_reason": finish_reason,
                    }
                ],
            }
            if self.cache:
                if self.verbose:
                    print("Llama._create_completion: cache save", file=sys.stderr)
                self.cache[prompt_tokens + completion_tokens] = self.save_state()
                print("Llama._create_completion: cache saved", file=sys.stderr)
            return

        if self.cache:
            if self.verbose:
                print("Llama._create_completion: cache save", file=sys.stderr)
            self.cache[prompt_tokens + completion_tokens] = self.save_state()

        text_str = text.decode("utf-8", errors="ignore")

        if echo:
            text_str = prompt + text_str

        if suffix is not None:
            text_str = text_str + suffix

        logprobs_or_none: Optional[CompletionLogprobs] = None
        if logprobs is not None:
            text_offset = 0 if echo else len(prompt)
            token_offset = 0 if echo else len(prompt_tokens[1:])
            text_offsets: List[int] = []
            token_logprobs: List[Optional[float]] = []
            tokens: List[str] = []
            top_logprobs: List[Optional[Dict[str, float]]] = []

            if echo:
                # Remove leading BOS token
                all_tokens = prompt_tokens[1:] + completion_tokens
            else:
                all_tokens = completion_tokens

            all_token_strs = [
                self.detokenize([token]).decode("utf-8", errors="ignore")
                for token in all_tokens
            ]
            all_logprobs = [
                Llama.logits_to_logprobs(row.tolist()) for row in self._scores
            ][token_offset:]
            for token, token_str, logprobs_token in zip(
                all_tokens, all_token_strs, all_logprobs
            ):
                if token == self.token_bos():
                    continue
                text_offsets.append(text_offset)
                text_offset += len(token_str)
                tokens.append(token_str)
                sorted_logprobs = list(
                    sorted(
                        zip(logprobs_token, range(len(logprobs_token))), reverse=True
                    )
                )
                token_logprobs.append(logprobs_token[int(token)])
                top_logprob: Optional[Dict[str, float]] = {
                    self.detokenize([i]).decode("utf-8", errors="ignore"): logprob
                    for logprob, i in sorted_logprobs[:logprobs]
                }
                top_logprob.update({token_str: logprobs_token[int(token)]})
                top_logprobs.append(top_logprob)
            # Weird idosincracy of the OpenAI API where
            # token_logprobs and top_logprobs are null for
            # the first token.
            if echo and len(all_tokens) > 0:
                token_logprobs[0] = None
                top_logprobs[0] = None
            logprobs_or_none = {
                "tokens": tokens,
                "text_offset": text_offsets,
                "token_logprobs": token_logprobs,
                "top_logprobs": top_logprobs,
            }

        yield {
            "id": completion_id,
            "object": "text_completion",
            "created": created,
            "model": model_name,
            "choices": [
                {
                    "text": text_str,
                    "index": 0,
                    "logprobs": logprobs_or_none,
                    "finish_reason": finish_reason,
                }
            ],
            "usage": {
                "prompt_tokens": len(prompt_tokens),
                "completion_tokens": len(completion_tokens),
                "total_tokens": len(prompt_tokens) + len(completion_tokens),
            },
        }

    def create_completion(
        self,
        prompt: Union[str, List[int]],
        suffix: Optional[str] = None,
        max_tokens: Optional[int] = 16,
        temperature: float = 0.8,
        top_p: float = 0.95,
        min_p: float = 0.05,
        typical_p: float = 1.0,
        logprobs: Optional[int] = None,
        echo: bool = False,
        stop: Optional[Union[str, List[str]]] = [],
        frequency_penalty: float = 0.0,
        presence_penalty: float = 0.0,
        repeat_penalty: float = 1.1,
        top_k: int = 40,
        stream: bool = False,
        seed: Optional[int] = None,
        tfs_z: float = 1.0,
        mirostat_mode: int = 0,
        mirostat_tau: float = 5.0,
        mirostat_eta: float = 0.1,
        model: Optional[str] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        logits_processor: Optional[LogitsProcessorList] = None,
        grammar: Optional[LlamaGrammar] = None,
        logit_bias: Optional[Dict[str, float]] = None,
    ) -> Union[CreateCompletionResponse, Iterator[CreateCompletionStreamResponse]]:
        """Generate text from a prompt.

        Args:
            prompt: The prompt to generate text from.
            suffix: A suffix to append to the generated text. If None, no suffix is appended.
            max_tokens: The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
            min_p: The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
            typical_p: The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
            logprobs: The number of logprobs to return. If None, no logprobs are returned.
            echo: Whether to echo the prompt.
            stop: A list of strings to stop generation when encountered.
            frequency_penalty: The penalty to apply to tokens based on their frequency in the prompt.
            presence_penalty: The penalty to apply to tokens based on their presence in the prompt.
            repeat_penalty: The penalty to apply to repeated tokens.
            top_k: The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
            stream: Whether to stream the results.
            seed: The seed to use for sampling.
            tfs_z: The tail-free sampling parameter. Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
            mirostat_mode: The mirostat sampling mode.
            mirostat_tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
            mirostat_eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
            model: The name to use for the model in the completion object.
            stopping_criteria: A list of stopping criteria to use.
            logits_processor: A list of logits processors to use.
            grammar: A grammar to use for constrained sampling.
            logit_bias: A logit bias to use.

        Raises:
            ValueError: If the requested tokens exceed the context window.
            RuntimeError: If the prompt fails to tokenize or the model fails to evaluate the prompt.

        Returns:
            Response object containing the generated text.
        """
        completion_or_chunks = self._create_completion(
            prompt=prompt,
            suffix=suffix,
            max_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            min_p=min_p,
            typical_p=typical_p,
            logprobs=logprobs,
            echo=echo,
            stop=stop,
            frequency_penalty=frequency_penalty,
            presence_penalty=presence_penalty,
            repeat_penalty=repeat_penalty,
            top_k=top_k,
            stream=stream,
            seed=seed,
            tfs_z=tfs_z,
            mirostat_mode=mirostat_mode,
            mirostat_tau=mirostat_tau,
            mirostat_eta=mirostat_eta,
            model=model,
            stopping_criteria=stopping_criteria,
            logits_processor=logits_processor,
            grammar=grammar,
            logit_bias=logit_bias,
        )
        if stream:
            chunks: Iterator[CreateCompletionStreamResponse] = completion_or_chunks
            return chunks
        completion: Completion = next(completion_or_chunks)  # type: ignore
        return completion

    def __call__(
        self,
        prompt: str,
        suffix: Optional[str] = None,
        max_tokens: int = 128,
        temperature: float = 0.8,
        top_p: float = 0.95,
        min_p: float = 0.05,
        typical_p: float = 1.0,
        logprobs: Optional[int] = None,
        echo: bool = False,
        stop: Optional[Union[str, List[str]]] = [],
        frequency_penalty: float = 0.0,
        presence_penalty: float = 0.0,
        repeat_penalty: float = 1.1,
        top_k: int = 40,
        stream: bool = False,
        seed: Optional[int] = None,
        tfs_z: float = 1.0,
        mirostat_mode: int = 0,
        mirostat_tau: float = 5.0,
        mirostat_eta: float = 0.1,
        model: Optional[str] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        logits_processor: Optional[LogitsProcessorList] = None,
        grammar: Optional[LlamaGrammar] = None,
        logit_bias: Optional[Dict[str, float]] = None,
    ) -> Union[CreateCompletionResponse, Iterator[CreateCompletionStreamResponse]]:
        """Generate text from a prompt.

        Args:
            prompt: The prompt to generate text from.
            suffix: A suffix to append to the generated text. If None, no suffix is appended.
            max_tokens: The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
            min_p: The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
            typical_p: The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
            logprobs: The number of logprobs to return. If None, no logprobs are returned.
            echo: Whether to echo the prompt.
            stop: A list of strings to stop generation when encountered.
            frequency_penalty: The penalty to apply to tokens based on their frequency in the prompt.
            presence_penalty: The penalty to apply to tokens based on their presence in the prompt.
            repeat_penalty: The penalty to apply to repeated tokens.
            top_k: The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
            stream: Whether to stream the results.
            seed: The seed to use for sampling.
            tfs_z: The tail-free sampling parameter. Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
            mirostat_mode: The mirostat sampling mode.
            mirostat_tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
            mirostat_eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
            model: The name to use for the model in the completion object.
            stopping_criteria: A list of stopping criteria to use.
            logits_processor: A list of logits processors to use.
            grammar: A grammar to use for constrained sampling.
            logit_bias: A logit bias to use.

        Raises:
            ValueError: If the requested tokens exceed the context window.
            RuntimeError: If the prompt fails to tokenize or the model fails to evaluate the prompt.

        Returns:
            Response object containing the generated text.
        """
        return self.create_completion(
            prompt=prompt,
            suffix=suffix,
            max_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            min_p=min_p,
            typical_p=typical_p,
            logprobs=logprobs,
            echo=echo,
            stop=stop,
            frequency_penalty=frequency_penalty,
            presence_penalty=presence_penalty,
            repeat_penalty=repeat_penalty,
            top_k=top_k,
            stream=stream,
            seed=seed,
            tfs_z=tfs_z,
            mirostat_mode=mirostat_mode,
            mirostat_tau=mirostat_tau,
            mirostat_eta=mirostat_eta,
            model=model,
            stopping_criteria=stopping_criteria,
            logits_processor=logits_processor,
            grammar=grammar,
            logit_bias=logit_bias,
        )

    def create_chat_completion(
        self,
        messages: List[ChatCompletionRequestMessage],
        functions: Optional[List[ChatCompletionFunction]] = None,
        function_call: Optional[ChatCompletionRequestFunctionCall] = None,
        tools: Optional[List[ChatCompletionTool]] = None,
        tool_choice: Optional[ChatCompletionToolChoiceOption] = None,
        temperature: float = 0.2,
        top_p: float = 0.95,
        top_k: int = 40,
        min_p: float = 0.05,
        typical_p: float = 1.0,
        stream: bool = False,
        stop: Optional[Union[str, List[str]]] = [],
        seed: Optional[int] = None,
        response_format: Optional[ChatCompletionRequestResponseFormat] = None,
        max_tokens: Optional[int] = None,
        presence_penalty: float = 0.0,
        frequency_penalty: float = 0.0,
        repeat_penalty: float = 1.1,
        tfs_z: float = 1.0,
        mirostat_mode: int = 0,
        mirostat_tau: float = 5.0,
        mirostat_eta: float = 0.1,
        model: Optional[str] = None,
        logits_processor: Optional[LogitsProcessorList] = None,
        grammar: Optional[LlamaGrammar] = None,
        logit_bias: Optional[Dict[str, float]] = None,
    ) -> Union[
        CreateChatCompletionResponse, Iterator[CreateChatCompletionStreamResponse]
    ]:
        """Generate a chat completion from a list of messages.

        Args:
            messages: A list of messages to generate a response for.
            functions: A list of functions to use for the chat completion.
            function_call: A function call to use for the chat completion.
            tools: A list of tools to use for the chat completion.
            tool_choice: A tool choice to use for the chat completion.
            temperature: The temperature to use for sampling.
            top_p: The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
            top_k: The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
            min_p: The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
            typical_p: The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
            stream: Whether to stream the results.
            stop: A list of strings to stop generation when encountered.
            seed: The seed to use for sampling.
            response_format: The response format to use for the chat completion. Use { "type": "json_object" } to contstrain output to only valid json.
            max_tokens: The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.
            presence_penalty: The penalty to apply to tokens based on their presence in the prompt.
            frequency_penalty: The penalty to apply to tokens based on their frequency in the prompt.
            repeat_penalty: The penalty to apply to repeated tokens.
            tfs_z: The tail-free sampling parameter.
            mirostat_mode: The mirostat sampling mode.
            mirostat_tau: The mirostat sampling tau parameter.
            mirostat_eta: The mirostat sampling eta parameter.
            model: The name to use for the model in the completion object.
            logits_processor: A list of logits processors to use.
            grammar: A grammar to use.
            logit_bias: A logit bias to use.

        Returns:
            Generated chat completion or a stream of chat completion chunks.
        """
        handler = self.chat_handler or llama_chat_format.get_chat_completion_handler(
            self.chat_format
        )
        return handler(
            llama=self,
            messages=messages,
            functions=functions,
            function_call=function_call,
            tools=tools,
            tool_choice=tool_choice,
            temperature=temperature,
            top_p=top_p,
            top_k=top_k,
            min_p=min_p,
            typical_p=typical_p,
            stream=stream,
            stop=stop,
            seed=seed,
            response_format=response_format,
            max_tokens=max_tokens,
            presence_penalty=presence_penalty,
            frequency_penalty=frequency_penalty,
            repeat_penalty=repeat_penalty,
            tfs_z=tfs_z,
            mirostat_mode=mirostat_mode,
            mirostat_tau=mirostat_tau,
            mirostat_eta=mirostat_eta,
            model=model,
            logits_processor=logits_processor,
            grammar=grammar,
            logit_bias=logit_bias,
        )

    def __getstate__(self):
        return dict(
            model_path=self.model_path,
            # Model Params
            n_gpu_layers=self.model_params.n_gpu_layers,
            main_gpu=self.model_params.main_gpu,
            tensor_split=self.tensor_split,
            vocab_only=self.model_params.vocab_only,
            use_mmap=self.model_params.use_mmap,
            use_mlock=self.model_params.use_mlock,
            # Context Params
            seed=self.context_params.seed,
            n_ctx=self.context_params.n_ctx,
            n_batch=self.n_batch,
            n_threads=self.context_params.n_threads,
            n_threads_batch=self.context_params.n_threads_batch,
            rope_scaling_type=self.context_params.rope_scaling_type,
            rope_freq_base=self.context_params.rope_freq_base,
            rope_freq_scale=self.context_params.rope_freq_scale,
            yarn_ext_factor=self.context_params.yarn_ext_factor,
            yarn_attn_factor=self.context_params.yarn_attn_factor,
            yarn_beta_fast=self.context_params.yarn_beta_fast,
            yarn_beta_slow=self.context_params.yarn_beta_slow,
            yarn_orig_ctx=self.context_params.yarn_orig_ctx,
            mul_mat_q=self.context_params.mul_mat_q,
            f16_kv=self.context_params.f16_kv,
            logits_all=self.context_params.logits_all,
            embedding=self.context_params.embedding,
            # Sampling Params
            last_n_tokens_size=self.last_n_tokens_size,
            # LoRA Params
            lora_base=self.lora_base,
            lora_scale=self.lora_scale,
            lora_path=self.lora_path,
            # Backend Params
            numa=self.numa,
            # Chat Format Params
            chat_format=self.chat_format,
            chat_handler=self.chat_handler,
            # Misc
            verbose=self.verbose,
        )

    def __setstate__(self, state):
        self.__init__(
            model_path=state["model_path"],
            # Model Params
            n_gpu_layers=state["n_gpu_layers"],
            main_gpu=state["main_gpu"],
            tensor_split=state["tensor_split"],
            vocab_only=state["vocab_only"],
            use_mmap=state["use_mmap"],
            use_mlock=state["use_mlock"],
            # Context Params
            seed=state["seed"],
            n_ctx=state["n_ctx"],
            n_batch=state["n_batch"],
            n_threads=state["n_threads"],
            n_threads_batch=state["n_threads_batch"],
            rope_freq_base=state["rope_freq_base"],
            rope_freq_scale=state["rope_freq_scale"],
            rope_scaling_type=state["rope_scaling_type"],
            yarn_ext_factor=state["yarn_ext_factor"],
            yarn_attn_factor=state["yarn_attn_factor"],
            yarn_beta_fast=state["yarn_beta_fast"],
            yarn_beta_slow=state["yarn_beta_slow"],
            yarn_orig_ctx=state["yarn_orig_ctx"],
            mul_mat_q=state["mul_mat_q"],
            f16_kv=state["f16_kv"],
            logits_all=state["logits_all"],
            embedding=state["embedding"],
            # Sampling Params
            last_n_tokens_size=state["last_n_tokens_size"],
            # LoRA Params
            lora_base=state["lora_base"],
            lora_path=state["lora_path"],
            # Backend Params
            numa=state["numa"],
            # Chat Format Params
            chat_format=state["chat_format"],
            chat_handler=state["chat_handler"],
            # Misc
            verbose=state["verbose"],
        )

    def save_state(self) -> LlamaState:
        assert self._ctx.ctx is not None
        if self.verbose:
            print("Llama.save_state: saving llama state", file=sys.stderr)
        state_size = llama_cpp.llama_get_state_size(self._ctx.ctx)
        if self.verbose:
            print(f"Llama.save_state: got state size: {state_size}", file=sys.stderr)
        llama_state = (llama_cpp.c_uint8 * int(state_size))()
        if self.verbose:
            print("Llama.save_state: allocated state", file=sys.stderr)
        n_bytes = llama_cpp.llama_copy_state_data(self._ctx.ctx, llama_state)
        if self.verbose:
            print(f"Llama.save_state: copied llama state: {n_bytes}", file=sys.stderr)
        if int(n_bytes) > int(state_size):
            raise RuntimeError("Failed to copy llama state data")
        llama_state_compact = (llama_cpp.c_uint8 * int(n_bytes))()
        llama_cpp.ctypes.memmove(llama_state_compact, llama_state, int(n_bytes))
        if self.verbose:
            print(
                f"Llama.save_state: saving {n_bytes} bytes of llama state",
                file=sys.stderr,
            )
        return LlamaState(
            scores=self.scores.copy(),
            input_ids=self.input_ids.copy(),
            n_tokens=self.n_tokens,
            llama_state=bytes(llama_state_compact),
            llama_state_size=n_bytes,
        )

    def load_state(self, state: LlamaState) -> None:
        assert self._ctx.ctx is not None
        self.scores = state.scores.copy()
        self.input_ids = state.input_ids.copy()
        self.n_tokens = state.n_tokens
        state_size = state.llama_state_size
        LLamaStateArrayType = llama_cpp.c_uint8 * state_size
        llama_state = LLamaStateArrayType.from_buffer_copy(state.llama_state)

        if llama_cpp.llama_set_state_data(self._ctx.ctx, llama_state) != state_size:
            raise RuntimeError("Failed to set llama state data")

    def n_ctx(self) -> int:
        """Return the context window size."""
        return self._ctx.n_ctx()

    def n_embd(self) -> int:
        """Return the embedding size."""
        return self._model.n_embd()

    def n_vocab(self) -> int:
        """Return the vocabulary size."""
        return self._model.n_vocab()

    def tokenizer(self) -> "LlamaTokenizer":
        """Return the tokenizer for this model."""
        return LlamaTokenizer(self)

    def token_eos(self) -> int:
        """Return the end-of-sequence token."""
        return self._model.token_eos()

    def token_bos(self) -> int:
        """Return the beginning-of-sequence token."""
        return self._model.token_bos()

    def token_nl(self) -> int:
        """Return the newline token."""
        return self._model.token_nl()

    @staticmethod
    def logits_to_logprobs(logits: List[float]) -> List[float]:
        exps = [math.exp(float(x)) for x in logits]
        sum_exps = sum(exps)
        return [math.log(x / sum_exps) for x in exps]

    @staticmethod
    def longest_token_prefix(a: Sequence[int], b: Sequence[int]):
        longest_prefix = 0
        for _a, _b in zip(a, b):
            if _a == _b:
                longest_prefix += 1
            else:
                break
        return longest_prefix

__init__(model_path, *, n_gpu_layers=0, main_gpu=0, tensor_split=None, vocab_only=False, use_mmap=True, use_mlock=False, seed=llama_cpp.LLAMA_DEFAULT_SEED, n_ctx=512, n_batch=512, n_threads=None, n_threads_batch=None, rope_scaling_type=llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED, rope_freq_base=0.0, rope_freq_scale=0.0, yarn_ext_factor=-1.0, yarn_attn_factor=1.0, yarn_beta_fast=32.0, yarn_beta_slow=1.0, yarn_orig_ctx=0, mul_mat_q=True, f16_kv=True, logits_all=False, embedding=False, last_n_tokens_size=64, lora_base=None, lora_scale=1.0, lora_path=None, numa=False, chat_format='llama-2', chat_handler=None, verbose=True, **kwargs)

Load a llama.cpp model from model_path.

Examples:

Basic usage

>>> import llama_cpp
>>> model = llama_cpp.Llama(
...     model_path="path/to/model",
... )
>>> print(model("The quick brown fox jumps ", stop=["."])["choices"][0]["text"])
the lazy dog

Loading a chat model

>>> import llama_cpp
>>> model = llama_cpp.Llama(
...     model_path="path/to/model",
...     chat_format="llama-2",
... )
>>> print(model.create_chat_completion(
...     messages=[{
...         "role": "user",
...         "content": "what is the meaning of life?"
...     }]
... ))

Parameters:

  • model_path (str) –

    Path to the model.

  • n_gpu_layers (int, default: 0 ) –

    Number of layers to offload to GPU (-ngl). If -1, all layers are offloaded.

  • main_gpu (int, default: 0 ) –

    The GPU that is used for scratch and small tensors.

  • tensor_split (Optional[List[float]], default: None ) –

    How split tensors should be distributed across GPUs. If None, the model is not split.

  • vocab_only (bool, default: False ) –

    Only load the vocabulary no weights.

  • use_mmap (bool, default: True ) –

    Use mmap if possible.

  • use_mlock (bool, default: False ) –

    Force the system to keep the model in RAM.

  • seed (int, default: LLAMA_DEFAULT_SEED ) –

    RNG seed, -1 for random

  • n_ctx (int, default: 512 ) –

    Text context, 0 = from model

  • n_batch (int, default: 512 ) –

    Prompt processing maximum batch size

  • n_threads (Optional[int], default: None ) –

    Number of threads to use for generation

  • n_threads_batch (Optional[int], default: None ) –

    Number of threads to use for batch processing

  • rope_scaling_type (Optional[int], default: LLAMA_ROPE_SCALING_UNSPECIFIED ) –

    RoPE scaling type, from enum llama_rope_scaling_type. ref: https://github.com/ggerganov/llama.cpp/pull/2054

  • rope_freq_base (float, default: 0.0 ) –

    RoPE base frequency, 0 = from model

  • rope_freq_scale (float, default: 0.0 ) –

    RoPE frequency scaling factor, 0 = from model

  • yarn_ext_factor (float, default: -1.0 ) –

    YaRN extrapolation mix factor, negative = from model

  • yarn_attn_factor (float, default: 1.0 ) –

    YaRN magnitude scaling factor

  • yarn_beta_fast (float, default: 32.0 ) –

    YaRN low correction dim

  • yarn_beta_slow (float, default: 1.0 ) –

    YaRN high correction dim

  • yarn_orig_ctx (int, default: 0 ) –

    YaRN original context size

  • f16_kv (bool, default: True ) –

    Use fp16 for KV cache, fp32 otherwise

  • logits_all (bool, default: False ) –

    Return logits for all tokens, not just the last token. Must be True for completion to return logprobs.

  • embedding (bool, default: False ) –

    Embedding mode only.

  • last_n_tokens_size (int, default: 64 ) –

    Maximum number of tokens to keep in the last_n_tokens deque.

  • lora_base (Optional[str], default: None ) –

    Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model.

  • lora_path (Optional[str], default: None ) –

    Path to a LoRA file to apply to the model.

  • numa (bool, default: False ) –

    Enable NUMA support. (NOTE: The initial value of this parameter is used for the remainder of the program as this value is set in llama_backend_init)

  • chat_format (str, default: 'llama-2' ) –

    String specifying the chat format to use when calling create_chat_completion.

  • chat_handler (Optional[LlamaChatCompletionHandler], default: None ) –

    Optional chat handler to use when calling create_chat_completion.

  • verbose (bool, default: True ) –

    Print verbose output to stderr.

Raises:

  • ValueError

    If the model path does not exist.

Returns:

  • A Llama instance.

Source code in llama_cpp/llama.py
def __init__(
    self,
    model_path: str,
    *,
    # Model Params
    n_gpu_layers: int = 0,
    main_gpu: int = 0,
    tensor_split: Optional[List[float]] = None,
    vocab_only: bool = False,
    use_mmap: bool = True,
    use_mlock: bool = False,
    # Context Params
    seed: int = llama_cpp.LLAMA_DEFAULT_SEED,
    n_ctx: int = 512,
    n_batch: int = 512,
    n_threads: Optional[int] = None,
    n_threads_batch: Optional[int] = None,
    rope_scaling_type: Optional[int] = llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED,
    rope_freq_base: float = 0.0,
    rope_freq_scale: float = 0.0,
    yarn_ext_factor: float = -1.0,
    yarn_attn_factor: float = 1.0,
    yarn_beta_fast: float = 32.0,
    yarn_beta_slow: float = 1.0,
    yarn_orig_ctx: int = 0,
    mul_mat_q: bool = True,
    f16_kv: bool = True,
    logits_all: bool = False,
    embedding: bool = False,
    # Sampling Params
    last_n_tokens_size: int = 64,
    # LoRA Params
    lora_base: Optional[str] = None,
    lora_scale: float = 1.0,
    lora_path: Optional[str] = None,
    # Backend Params
    numa: bool = False,
    # Chat Format Params
    chat_format: str = "llama-2",
    chat_handler: Optional[llama_chat_format.LlamaChatCompletionHandler] = None,
    # Misc
    verbose: bool = True,
    # Extra Params
    **kwargs,  # type: ignore
):
    """Load a llama.cpp model from `model_path`.

    Examples:
        Basic usage

        >>> import llama_cpp
        >>> model = llama_cpp.Llama(
        ...     model_path="path/to/model",
        ... )
        >>> print(model("The quick brown fox jumps ", stop=["."])["choices"][0]["text"])
        the lazy dog

        Loading a chat model

        >>> import llama_cpp
        >>> model = llama_cpp.Llama(
        ...     model_path="path/to/model",
        ...     chat_format="llama-2",
        ... )
        >>> print(model.create_chat_completion(
        ...     messages=[{
        ...         "role": "user",
        ...         "content": "what is the meaning of life?"
        ...     }]
        ... ))

    Args:
        model_path: Path to the model.
        n_gpu_layers: Number of layers to offload to GPU (-ngl). If -1, all layers are offloaded.
        main_gpu: The GPU that is used for scratch and small tensors.
        tensor_split: How split tensors should be distributed across GPUs. If None, the model is not split.
        vocab_only: Only load the vocabulary no weights.
        use_mmap: Use mmap if possible.
        use_mlock: Force the system to keep the model in RAM.
        seed: RNG seed, -1 for random
        n_ctx: Text context, 0 = from model
        n_batch: Prompt processing maximum batch size
        n_threads: Number of threads to use for generation
        n_threads_batch: Number of threads to use for batch processing
        rope_scaling_type: RoPE scaling type, from `enum llama_rope_scaling_type`. ref: https://github.com/ggerganov/llama.cpp/pull/2054
        rope_freq_base: RoPE base frequency, 0 = from model
        rope_freq_scale: RoPE frequency scaling factor, 0 = from model
        yarn_ext_factor: YaRN extrapolation mix factor, negative = from model
        yarn_attn_factor: YaRN magnitude scaling factor
        yarn_beta_fast: YaRN low correction dim
        yarn_beta_slow: YaRN high correction dim
        yarn_orig_ctx: YaRN original context size
        f16_kv: Use fp16 for KV cache, fp32 otherwise
        logits_all: Return logits for all tokens, not just the last token. Must be True for completion to return logprobs.
        embedding: Embedding mode only.
        last_n_tokens_size: Maximum number of tokens to keep in the last_n_tokens deque.
        lora_base: Optional path to base model, useful if using a quantized base model and you want to apply LoRA to an f16 model.
        lora_path: Path to a LoRA file to apply to the model.
        numa: Enable NUMA support. (NOTE: The initial value of this parameter is used for the remainder of the program as this value is set in llama_backend_init)
        chat_format: String specifying the chat format to use when calling create_chat_completion.
        chat_handler: Optional chat handler to use when calling create_chat_completion.
        verbose: Print verbose output to stderr.

    Raises:
        ValueError: If the model path does not exist.

    Returns:
        A Llama instance.
    """
    self.verbose = verbose

    self.numa = numa
    if not Llama.__backend_initialized:
        with suppress_stdout_stderr(disable=self.verbose):
            llama_cpp.llama_backend_init(self.numa)
        Llama.__backend_initialized = True

    self.model_path = model_path

    # Model Params
    self.model_params = llama_cpp.llama_model_default_params()
    self.model_params.n_gpu_layers = (
        0x7FFFFFFF if n_gpu_layers == -1 else n_gpu_layers
    )  # 0x7FFFFFFF is INT32 max, will be auto set to all layers
    self.model_params.main_gpu = main_gpu
    self.tensor_split = tensor_split
    self._p_tensor_split = None
    if self.tensor_split is not None:
        if len(self.tensor_split) > llama_cpp.LLAMA_MAX_DEVICES:
            raise ValueError(
                f"Attempt to split tensors that exceed maximum supported devices. Current LLAMA_MAX_DEVICES={llama_cpp.LLAMA_MAX_DEVICES}"
            )
        # Type conversion and expand the list to the length of LLAMA_MAX_DEVICES
        FloatArray = ctypes.c_float * llama_cpp.LLAMA_MAX_DEVICES
        self._c_tensor_split = FloatArray(
            *tensor_split  # type: ignore
        )  # keep a reference to the array so it is not gc'd
        self.model_params.tensor_split = self._c_tensor_split
    self.model_params.vocab_only = vocab_only
    self.model_params.use_mmap = use_mmap if lora_path is None else False
    self.model_params.use_mlock = use_mlock

    self.n_batch = min(n_ctx, n_batch)  # ???
    self.n_threads = n_threads or max(multiprocessing.cpu_count() // 2, 1)
    self.n_threads_batch = n_threads_batch or max(
        multiprocessing.cpu_count() // 2, 1
    )
    # Context Params
    self.context_params = llama_cpp.llama_context_default_params()
    self.context_params.seed = seed
    self.context_params.n_ctx = n_ctx
    self.context_params.n_batch = self.n_batch
    self.context_params.n_threads = self.n_threads
    self.context_params.n_threads_batch = self.n_threads_batch
    self.context_params.rope_scaling_type = (
        rope_scaling_type
        if rope_scaling_type is not None
        else llama_cpp.LLAMA_ROPE_SCALING_UNSPECIFIED
    )
    self.context_params.rope_freq_base = (
        rope_freq_base if rope_freq_base != 0.0 else 0
    )
    self.context_params.rope_freq_scale = (
        rope_freq_scale if rope_freq_scale != 0.0 else 0
    )
    self.context_params.yarn_ext_factor = (
        yarn_ext_factor if yarn_ext_factor != 0.0 else 0
    )
    self.context_params.yarn_attn_factor = (
        yarn_attn_factor if yarn_attn_factor != 0.0 else 0
    )
    self.context_params.yarn_beta_fast = (
        yarn_beta_fast if yarn_beta_fast != 0.0 else 0
    )
    self.context_params.yarn_beta_slow = (
        yarn_beta_slow if yarn_beta_slow != 0.0 else 0
    )
    self.context_params.yarn_orig_ctx = yarn_orig_ctx if yarn_orig_ctx != 0 else 0
    self.context_params.mul_mat_q = mul_mat_q
    self.context_params.f16_kv = f16_kv
    self.context_params.logits_all = logits_all
    self.context_params.embedding = embedding

    # Sampling Params
    self.last_n_tokens_size = last_n_tokens_size

    self.cache: Optional[BaseLlamaCache] = None

    self.lora_base = lora_base
    self.lora_scale = lora_scale
    self.lora_path = lora_path

    if not os.path.exists(model_path):
        raise ValueError(f"Model path does not exist: {model_path}")

    self._model = _LlamaModel(
        path_model=self.model_path, params=self.model_params, verbose=self.verbose
    )

    self._ctx = _LlamaContext(
        model=self._model,
        params=self.context_params,
        verbose=self.verbose,
    )

    self._batch = _LlamaBatch(
        n_tokens=self.n_batch,
        embd=0,
        n_seq_max=self.context_params.n_ctx,
        verbose=self.verbose,
    )

    if self.lora_path:
        if self._model.apply_lora_from_file(
            self.lora_path,
            self.lora_scale,
            self.lora_base,
            self.n_threads,
        ):
            raise RuntimeError(
                f"Failed to apply LoRA from lora path: {self.lora_path} to base path: {self.lora_base}"
            )

    if self.verbose:
        print(llama_cpp.llama_print_system_info().decode("utf-8"), file=sys.stderr)

    self.chat_format = chat_format
    self.chat_handler = chat_handler

    self._n_vocab = self.n_vocab()
    self._n_ctx = self.n_ctx()

    self._token_nl = self.token_nl()
    self._token_eos = self.token_eos()

    self._candidates = _LlamaTokenDataArray(n_vocab=self._n_vocab)

    self.n_tokens = 0
    self.input_ids: npt.NDArray[np.intc] = np.ndarray((n_ctx,), dtype=np.intc)
    self.scores: npt.NDArray[np.single] = np.ndarray(
        (n_ctx, self._n_vocab), dtype=np.single
    )

tokenize(text, add_bos=True, special=False)

Tokenize a string.

Parameters:

  • text (bytes) –

    The utf-8 encoded string to tokenize.

Raises:

Returns:

Source code in llama_cpp/llama.py
def tokenize(
    self, text: bytes, add_bos: bool = True, special: bool = False
) -> List[int]:
    """Tokenize a string.

    Args:
        text: The utf-8 encoded string to tokenize.

    Raises:
        RuntimeError: If the tokenization failed.

    Returns:
        A list of tokens.
    """
    return self._model.tokenize(text, add_bos, special)

detokenize(tokens)

Detokenize a list of tokens.

Parameters:

  • tokens (List[int]) –

    The list of tokens to detokenize.

Returns:

  • bytes

    The detokenized string.

Source code in llama_cpp/llama.py
def detokenize(self, tokens: List[int]) -> bytes:
    """Detokenize a list of tokens.

    Args:
        tokens: The list of tokens to detokenize.

    Returns:
        The detokenized string.
    """
    return self._model.detokenize(tokens)

reset()

Reset the model state.

Source code in llama_cpp/llama.py
def reset(self):
    """Reset the model state."""
    self.n_tokens = 0

eval(tokens)

Evaluate a list of tokens.

Parameters:

  • tokens (Sequence[int]) –

    The list of tokens to evaluate.

Source code in llama_cpp/llama.py
def eval(self, tokens: Sequence[int]):
    """Evaluate a list of tokens.

    Args:
        tokens: The list of tokens to evaluate.
    """
    assert self._ctx.ctx is not None
    assert self._batch.batch is not None
    self._ctx.kv_cache_seq_rm(-1, self.n_tokens, -1)
    for i in range(0, len(tokens), self.n_batch):
        batch = tokens[i : min(len(tokens), i + self.n_batch)]
        n_past = self.n_tokens
        n_tokens = len(batch)
        self._batch.set_batch(
            batch=batch, n_past=n_past, logits_all=self.context_params.logits_all
        )
        self._ctx.decode(self._batch)
        # Save tokens
        self.input_ids[n_past : n_past + n_tokens] = batch
        # Save logits
        rows = n_tokens
        cols = self._n_vocab
        offset = (
            0 if self.context_params.logits_all else n_tokens - 1
        )  # NOTE: Only save the last token logits if logits_all is False
        self.scores[n_past + offset : n_past + n_tokens, :].reshape(-1)[
            :
        ] = self._ctx.get_logits()[offset * cols : rows * cols]
        # Update n_tokens
        self.n_tokens += n_tokens

sample(top_k=40, top_p=0.95, min_p=0.05, typical_p=1.0, temp=0.8, repeat_penalty=1.1, frequency_penalty=0.0, presence_penalty=0.0, tfs_z=1.0, mirostat_mode=0, mirostat_eta=0.1, mirostat_tau=5.0, penalize_nl=True, logits_processor=None, grammar=None)

Sample a token from the model.

Parameters:

  • top_k (int, default: 40 ) –

    The top-k sampling parameter.

  • top_p (float, default: 0.95 ) –

    The top-p sampling parameter.

  • temp (float, default: 0.8 ) –

    The temperature parameter.

  • repeat_penalty (float, default: 1.1 ) –

    The repeat penalty parameter.

Returns:

  • The sampled token.

Source code in llama_cpp/llama.py
def sample(
    self,
    top_k: int = 40,
    top_p: float = 0.95,
    min_p: float = 0.05,
    typical_p: float = 1.0,
    temp: float = 0.80,
    repeat_penalty: float = 1.1,
    frequency_penalty: float = 0.0,
    presence_penalty: float = 0.0,
    tfs_z: float = 1.0,
    mirostat_mode: int = 0,
    mirostat_eta: float = 0.1,
    mirostat_tau: float = 5.0,
    penalize_nl: bool = True,
    logits_processor: Optional[LogitsProcessorList] = None,
    grammar: Optional[LlamaGrammar] = None,
):
    """Sample a token from the model.

    Args:
        top_k: The top-k sampling parameter.
        top_p: The top-p sampling parameter.
        temp: The temperature parameter.
        repeat_penalty: The repeat penalty parameter.

    Returns:
        The sampled token.
    """
    assert self._ctx is not None
    assert self.n_tokens > 0
    last_n_tokens_data = [llama_cpp.llama_token(0)] * max(
        0, self.last_n_tokens_size - self.n_tokens
    ) + self._input_ids[-self.last_n_tokens_size :].tolist()
    last_n_tokens_size = len(last_n_tokens_data)
    n_vocab = self._n_vocab
    n_ctx = self._n_ctx
    top_k = n_vocab if top_k <= 0 else top_k
    last_n_tokens_size = n_ctx if last_n_tokens_size < 0 else last_n_tokens_size
    last_n_tokens_data_c = (llama_cpp.llama_token * last_n_tokens_size)(
        *last_n_tokens_data
    )
    logits: npt.NDArray[np.single] = self._scores[-1, :]

    if logits_processor is not None:
        logits[:] = logits_processor(self._input_ids, logits)

    nl_logit = logits[self._token_nl]
    self._candidates.copy_logits(logits)
    self._ctx.sample_repetition_penalties(
        candidates=self._candidates,
        last_tokens_data=last_n_tokens_data_c,
        penalty_last_n=last_n_tokens_size,
        penalty_repeat=repeat_penalty,
        penalty_freq=frequency_penalty,
        penalty_present=presence_penalty,
    )
    if not penalize_nl:
        self._candidates.candidates.data[self._token_nl].logit = llama_cpp.c_float(
            nl_logit
        )

    if grammar is not None:
        self._ctx.sample_grammar(
            candidates=self._candidates,
            grammar=grammar,
        )

    if temp < 0.0:
        self._ctx.sample_softmax(candidates=self._candidates)
        id = self._candidates.candidates.data[0].id
    elif temp == 0.0:
        id = self._ctx.sample_token_greedy(candidates=self._candidates)
    elif mirostat_mode == 1:
        self._ctx.sample_temp(candidates=self._candidates, temp=temp)
        id = self._ctx.sample_token_mirostat(
            candidates=self._candidates,
            tau=mirostat_tau,
            eta=mirostat_eta,
            mu=2.0 * mirostat_tau,
            m=100,
        )
    elif mirostat_mode == 2:
        self._ctx.sample_temp(candidates=self._candidates, temp=temp)
        id = self._ctx.sample_token_mirostat_v2(
            candidates=self._candidates,
            tau=mirostat_tau,
            eta=mirostat_eta,
            mu=2.0 * mirostat_tau,
        )
    else:
        self._ctx.sample_top_k(candidates=self._candidates, k=top_k, min_keep=1)
        self._ctx.sample_tail_free(candidates=self._candidates, z=tfs_z, min_keep=1)
        self._ctx.sample_typical(
            candidates=self._candidates, p=typical_p, min_keep=1
        )
        self._ctx.sample_top_p(candidates=self._candidates, p=top_p, min_keep=1)
        self._ctx.sample_min_p(candidates=self._candidates, p=min_p, min_keep=1)
        self._ctx.sample_temp(candidates=self._candidates, temp=temp)
        id = self._ctx.sample_token(candidates=self._candidates)
    if grammar is not None:
        self._ctx.grammar_accept_token(grammar=grammar, token=id)
    return id

generate(tokens, top_k=40, top_p=0.95, min_p=0.05, typical_p=1.0, temp=0.8, repeat_penalty=1.1, reset=True, frequency_penalty=0.0, presence_penalty=0.0, tfs_z=1.0, mirostat_mode=0, mirostat_tau=5.0, mirostat_eta=0.1, logits_processor=None, stopping_criteria=None, grammar=None)

Create a generator of tokens from a prompt.

Examples:

>>> llama = Llama("models/ggml-7b.bin")
>>> tokens = llama.tokenize(b"Hello, world!")
>>> for token in llama.generate(tokens, top_k=40, top_p=0.95, temp=1.0, repeat_penalty=1.1):
...     print(llama.detokenize([token]))

Parameters:

  • tokens (Sequence[int]) –

    The prompt tokens.

  • top_k (int, default: 40 ) –

    The top-k sampling parameter.

  • top_p (float, default: 0.95 ) –

    The top-p sampling parameter.

  • temp (float, default: 0.8 ) –

    The temperature parameter.

  • repeat_penalty (float, default: 1.1 ) –

    The repeat penalty parameter.

  • reset (bool, default: True ) –

    Whether to reset the model state.

Yields:

  • int

    The generated tokens.

Source code in llama_cpp/llama.py
def generate(
    self,
    tokens: Sequence[int],
    top_k: int = 40,
    top_p: float = 0.95,
    min_p: float = 0.05,
    typical_p: float = 1.0,
    temp: float = 0.80,
    repeat_penalty: float = 1.1,
    reset: bool = True,
    frequency_penalty: float = 0.0,
    presence_penalty: float = 0.0,
    tfs_z: float = 1.0,
    mirostat_mode: int = 0,
    mirostat_tau: float = 5.0,
    mirostat_eta: float = 0.1,
    logits_processor: Optional[LogitsProcessorList] = None,
    stopping_criteria: Optional[StoppingCriteriaList] = None,
    grammar: Optional[LlamaGrammar] = None,
) -> Generator[int, Optional[Sequence[int]], None]:
    """Create a generator of tokens from a prompt.

    Examples:
        >>> llama = Llama("models/ggml-7b.bin")
        >>> tokens = llama.tokenize(b"Hello, world!")
        >>> for token in llama.generate(tokens, top_k=40, top_p=0.95, temp=1.0, repeat_penalty=1.1):
        ...     print(llama.detokenize([token]))

    Args:
        tokens: The prompt tokens.
        top_k: The top-k sampling parameter.
        top_p: The top-p sampling parameter.
        temp: The temperature parameter.
        repeat_penalty: The repeat penalty parameter.
        reset: Whether to reset the model state.

    Yields:
        The generated tokens.
    """
    if reset and self.n_tokens > 0:
        longest_prefix = 0
        for a, b in zip(self._input_ids, tokens[:-1]):
            if a == b:
                longest_prefix += 1
            else:
                break
        if longest_prefix > 0:
            if self.verbose:
                print("Llama.generate: prefix-match hit", file=sys.stderr)
            reset = False
            tokens = tokens[longest_prefix:]
            self.n_tokens = longest_prefix

    if reset:
        self.reset()

    if grammar is not None:
        grammar.reset()

    while True:
        self.eval(tokens)
        token = self.sample(
            top_k=top_k,
            top_p=top_p,
            min_p=min_p,
            typical_p=typical_p,
            temp=temp,
            repeat_penalty=repeat_penalty,
            frequency_penalty=frequency_penalty,
            presence_penalty=presence_penalty,
            tfs_z=tfs_z,
            mirostat_mode=mirostat_mode,
            mirostat_tau=mirostat_tau,
            mirostat_eta=mirostat_eta,
            logits_processor=logits_processor,
            grammar=grammar,
        )
        if stopping_criteria is not None and stopping_criteria(
            self._input_ids, self._scores[-1, :]
        ):
            return
        tokens_or_none = yield token
        tokens = [token]
        if tokens_or_none is not None:
            tokens.extend(tokens_or_none)

create_embedding(input, model=None)

Embed a string.

Parameters:

Returns:

Source code in llama_cpp/llama.py
def create_embedding(
    self, input: Union[str, List[str]], model: Optional[str] = None
) -> CreateEmbeddingResponse:
    """Embed a string.

    Args:
        input: The utf-8 encoded string to embed.

    Returns:
        An embedding object.
    """
    assert self._ctx.ctx is not None
    assert self._model.model is not None
    model_name: str = model if model is not None else self.model_path

    if self.context_params.embedding == False:
        raise RuntimeError(
            "Llama model must be created with embedding=True to call this method"
        )

    if self.verbose:
        llama_cpp.llama_reset_timings(self._ctx.ctx)

    if isinstance(input, str):
        inputs = [input]
    else:
        inputs = input

    data: List[Embedding] = []
    total_tokens = 0
    for index, input in enumerate(inputs):
        tokens = self.tokenize(input.encode("utf-8"), special=True)
        self.reset()
        self.eval(tokens)
        n_tokens = len(tokens)
        total_tokens += n_tokens
        embedding = llama_cpp.llama_get_embeddings(self._ctx.ctx)[
            : llama_cpp.llama_n_embd(self._model.model)
        ]

        data.append(
            {
                "object": "embedding",
                "embedding": embedding,
                "index": index,
            }
        )
    if self.verbose:
        llama_cpp.llama_print_timings(self._ctx.ctx)

    return {
        "object": "list",
        "data": data,
        "model": model_name,
        "usage": {
            "prompt_tokens": total_tokens,
            "total_tokens": total_tokens,
        },
    }

embed(input)

Embed a string.

Parameters:

  • input (str) –

    The utf-8 encoded string to embed.

Returns:

Source code in llama_cpp/llama.py
def embed(self, input: str) -> List[float]:
    """Embed a string.

    Args:
        input: The utf-8 encoded string to embed.

    Returns:
        A list of embeddings
    """
    return list(map(float, self.create_embedding(input)["data"][0]["embedding"]))

create_completion(prompt, suffix=None, max_tokens=16, temperature=0.8, top_p=0.95, min_p=0.05, typical_p=1.0, logprobs=None, echo=False, stop=[], frequency_penalty=0.0, presence_penalty=0.0, repeat_penalty=1.1, top_k=40, stream=False, seed=None, tfs_z=1.0, mirostat_mode=0, mirostat_tau=5.0, mirostat_eta=0.1, model=None, stopping_criteria=None, logits_processor=None, grammar=None, logit_bias=None)

Generate text from a prompt.

Parameters:

  • prompt (Union[str, List[int]]) –

    The prompt to generate text from.

  • suffix (Optional[str], default: None ) –

    A suffix to append to the generated text. If None, no suffix is appended.

  • max_tokens (Optional[int], default: 16 ) –

    The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.

  • temperature (float, default: 0.8 ) –

    The temperature to use for sampling.

  • top_p (float, default: 0.95 ) –

    The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751

  • min_p (float, default: 0.05 ) –

    The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841

  • typical_p (float, default: 1.0 ) –

    The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.

  • logprobs (Optional[int], default: None ) –

    The number of logprobs to return. If None, no logprobs are returned.

  • echo (bool, default: False ) –

    Whether to echo the prompt.

  • stop (Optional[Union[str, List[str]]], default: [] ) –

    A list of strings to stop generation when encountered.

  • frequency_penalty (float, default: 0.0 ) –

    The penalty to apply to tokens based on their frequency in the prompt.

  • presence_penalty (float, default: 0.0 ) –

    The penalty to apply to tokens based on their presence in the prompt.

  • repeat_penalty (float, default: 1.1 ) –

    The penalty to apply to repeated tokens.

  • top_k (int, default: 40 ) –

    The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751

  • stream (bool, default: False ) –

    Whether to stream the results.

  • seed (Optional[int], default: None ) –

    The seed to use for sampling.

  • tfs_z (float, default: 1.0 ) –

    The tail-free sampling parameter. Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.

  • mirostat_mode (int, default: 0 ) –

    The mirostat sampling mode.

  • mirostat_tau (float, default: 5.0 ) –

    The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.

  • mirostat_eta (float, default: 0.1 ) –

    The learning rate used to update mu based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause mu to be updated more quickly, while a smaller learning rate will result in slower updates.

  • model (Optional[str], default: None ) –

    The name to use for the model in the completion object.

  • stopping_criteria (Optional[StoppingCriteriaList], default: None ) –

    A list of stopping criteria to use.

  • logits_processor (Optional[LogitsProcessorList], default: None ) –

    A list of logits processors to use.

  • grammar (Optional[LlamaGrammar], default: None ) –

    A grammar to use for constrained sampling.

  • logit_bias (Optional[Dict[str, float]], default: None ) –

    A logit bias to use.

Raises:

  • ValueError

    If the requested tokens exceed the context window.

  • RuntimeError

    If the prompt fails to tokenize or the model fails to evaluate the prompt.

Returns:

Source code in llama_cpp/llama.py
def create_completion(
    self,
    prompt: Union[str, List[int]],
    suffix: Optional[str] = None,
    max_tokens: Optional[int] = 16,
    temperature: float = 0.8,
    top_p: float = 0.95,
    min_p: float = 0.05,
    typical_p: float = 1.0,
    logprobs: Optional[int] = None,
    echo: bool = False,
    stop: Optional[Union[str, List[str]]] = [],
    frequency_penalty: float = 0.0,
    presence_penalty: float = 0.0,
    repeat_penalty: float = 1.1,
    top_k: int = 40,
    stream: bool = False,
    seed: Optional[int] = None,
    tfs_z: float = 1.0,
    mirostat_mode: int = 0,
    mirostat_tau: float = 5.0,
    mirostat_eta: float = 0.1,
    model: Optional[str] = None,
    stopping_criteria: Optional[StoppingCriteriaList] = None,
    logits_processor: Optional[LogitsProcessorList] = None,
    grammar: Optional[LlamaGrammar] = None,
    logit_bias: Optional[Dict[str, float]] = None,
) -> Union[CreateCompletionResponse, Iterator[CreateCompletionStreamResponse]]:
    """Generate text from a prompt.

    Args:
        prompt: The prompt to generate text from.
        suffix: A suffix to append to the generated text. If None, no suffix is appended.
        max_tokens: The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.
        temperature: The temperature to use for sampling.
        top_p: The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
        min_p: The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
        typical_p: The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
        logprobs: The number of logprobs to return. If None, no logprobs are returned.
        echo: Whether to echo the prompt.
        stop: A list of strings to stop generation when encountered.
        frequency_penalty: The penalty to apply to tokens based on their frequency in the prompt.
        presence_penalty: The penalty to apply to tokens based on their presence in the prompt.
        repeat_penalty: The penalty to apply to repeated tokens.
        top_k: The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
        stream: Whether to stream the results.
        seed: The seed to use for sampling.
        tfs_z: The tail-free sampling parameter. Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
        mirostat_mode: The mirostat sampling mode.
        mirostat_tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
        mirostat_eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
        model: The name to use for the model in the completion object.
        stopping_criteria: A list of stopping criteria to use.
        logits_processor: A list of logits processors to use.
        grammar: A grammar to use for constrained sampling.
        logit_bias: A logit bias to use.

    Raises:
        ValueError: If the requested tokens exceed the context window.
        RuntimeError: If the prompt fails to tokenize or the model fails to evaluate the prompt.

    Returns:
        Response object containing the generated text.
    """
    completion_or_chunks = self._create_completion(
        prompt=prompt,
        suffix=suffix,
        max_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        min_p=min_p,
        typical_p=typical_p,
        logprobs=logprobs,
        echo=echo,
        stop=stop,
        frequency_penalty=frequency_penalty,
        presence_penalty=presence_penalty,
        repeat_penalty=repeat_penalty,
        top_k=top_k,
        stream=stream,
        seed=seed,
        tfs_z=tfs_z,
        mirostat_mode=mirostat_mode,
        mirostat_tau=mirostat_tau,
        mirostat_eta=mirostat_eta,
        model=model,
        stopping_criteria=stopping_criteria,
        logits_processor=logits_processor,
        grammar=grammar,
        logit_bias=logit_bias,
    )
    if stream:
        chunks: Iterator[CreateCompletionStreamResponse] = completion_or_chunks
        return chunks
    completion: Completion = next(completion_or_chunks)  # type: ignore
    return completion

__call__(prompt, suffix=None, max_tokens=128, temperature=0.8, top_p=0.95, min_p=0.05, typical_p=1.0, logprobs=None, echo=False, stop=[], frequency_penalty=0.0, presence_penalty=0.0, repeat_penalty=1.1, top_k=40, stream=False, seed=None, tfs_z=1.0, mirostat_mode=0, mirostat_tau=5.0, mirostat_eta=0.1, model=None, stopping_criteria=None, logits_processor=None, grammar=None, logit_bias=None)

Generate text from a prompt.

Parameters:

  • prompt (str) –

    The prompt to generate text from.

  • suffix (Optional[str], default: None ) –

    A suffix to append to the generated text. If None, no suffix is appended.

  • max_tokens (int, default: 128 ) –

    The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.

  • temperature (float, default: 0.8 ) –

    The temperature to use for sampling.

  • top_p (float, default: 0.95 ) –

    The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751

  • min_p (float, default: 0.05 ) –

    The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841

  • typical_p (float, default: 1.0 ) –

    The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.

  • logprobs (Optional[int], default: None ) –

    The number of logprobs to return. If None, no logprobs are returned.

  • echo (bool, default: False ) –

    Whether to echo the prompt.

  • stop (Optional[Union[str, List[str]]], default: [] ) –

    A list of strings to stop generation when encountered.

  • frequency_penalty (float, default: 0.0 ) –

    The penalty to apply to tokens based on their frequency in the prompt.

  • presence_penalty (float, default: 0.0 ) –

    The penalty to apply to tokens based on their presence in the prompt.

  • repeat_penalty (float, default: 1.1 ) –

    The penalty to apply to repeated tokens.

  • top_k (int, default: 40 ) –

    The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751

  • stream (bool, default: False ) –

    Whether to stream the results.

  • seed (Optional[int], default: None ) –

    The seed to use for sampling.

  • tfs_z (float, default: 1.0 ) –

    The tail-free sampling parameter. Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.

  • mirostat_mode (int, default: 0 ) –

    The mirostat sampling mode.

  • mirostat_tau (float, default: 5.0 ) –

    The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.

  • mirostat_eta (float, default: 0.1 ) –

    The learning rate used to update mu based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause mu to be updated more quickly, while a smaller learning rate will result in slower updates.

  • model (Optional[str], default: None ) –

    The name to use for the model in the completion object.

  • stopping_criteria (Optional[StoppingCriteriaList], default: None ) –

    A list of stopping criteria to use.

  • logits_processor (Optional[LogitsProcessorList], default: None ) –

    A list of logits processors to use.

  • grammar (Optional[LlamaGrammar], default: None ) –

    A grammar to use for constrained sampling.

  • logit_bias (Optional[Dict[str, float]], default: None ) –

    A logit bias to use.

Raises:

  • ValueError

    If the requested tokens exceed the context window.

  • RuntimeError

    If the prompt fails to tokenize or the model fails to evaluate the prompt.

Returns:

Source code in llama_cpp/llama.py
def __call__(
    self,
    prompt: str,
    suffix: Optional[str] = None,
    max_tokens: int = 128,
    temperature: float = 0.8,
    top_p: float = 0.95,
    min_p: float = 0.05,
    typical_p: float = 1.0,
    logprobs: Optional[int] = None,
    echo: bool = False,
    stop: Optional[Union[str, List[str]]] = [],
    frequency_penalty: float = 0.0,
    presence_penalty: float = 0.0,
    repeat_penalty: float = 1.1,
    top_k: int = 40,
    stream: bool = False,
    seed: Optional[int] = None,
    tfs_z: float = 1.0,
    mirostat_mode: int = 0,
    mirostat_tau: float = 5.0,
    mirostat_eta: float = 0.1,
    model: Optional[str] = None,
    stopping_criteria: Optional[StoppingCriteriaList] = None,
    logits_processor: Optional[LogitsProcessorList] = None,
    grammar: Optional[LlamaGrammar] = None,
    logit_bias: Optional[Dict[str, float]] = None,
) -> Union[CreateCompletionResponse, Iterator[CreateCompletionStreamResponse]]:
    """Generate text from a prompt.

    Args:
        prompt: The prompt to generate text from.
        suffix: A suffix to append to the generated text. If None, no suffix is appended.
        max_tokens: The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.
        temperature: The temperature to use for sampling.
        top_p: The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
        min_p: The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
        typical_p: The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
        logprobs: The number of logprobs to return. If None, no logprobs are returned.
        echo: Whether to echo the prompt.
        stop: A list of strings to stop generation when encountered.
        frequency_penalty: The penalty to apply to tokens based on their frequency in the prompt.
        presence_penalty: The penalty to apply to tokens based on their presence in the prompt.
        repeat_penalty: The penalty to apply to repeated tokens.
        top_k: The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
        stream: Whether to stream the results.
        seed: The seed to use for sampling.
        tfs_z: The tail-free sampling parameter. Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
        mirostat_mode: The mirostat sampling mode.
        mirostat_tau: The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
        mirostat_eta: The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
        model: The name to use for the model in the completion object.
        stopping_criteria: A list of stopping criteria to use.
        logits_processor: A list of logits processors to use.
        grammar: A grammar to use for constrained sampling.
        logit_bias: A logit bias to use.

    Raises:
        ValueError: If the requested tokens exceed the context window.
        RuntimeError: If the prompt fails to tokenize or the model fails to evaluate the prompt.

    Returns:
        Response object containing the generated text.
    """
    return self.create_completion(
        prompt=prompt,
        suffix=suffix,
        max_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        min_p=min_p,
        typical_p=typical_p,
        logprobs=logprobs,
        echo=echo,
        stop=stop,
        frequency_penalty=frequency_penalty,
        presence_penalty=presence_penalty,
        repeat_penalty=repeat_penalty,
        top_k=top_k,
        stream=stream,
        seed=seed,
        tfs_z=tfs_z,
        mirostat_mode=mirostat_mode,
        mirostat_tau=mirostat_tau,
        mirostat_eta=mirostat_eta,
        model=model,
        stopping_criteria=stopping_criteria,
        logits_processor=logits_processor,
        grammar=grammar,
        logit_bias=logit_bias,
    )

create_chat_completion(messages, functions=None, function_call=None, tools=None, tool_choice=None, temperature=0.2, top_p=0.95, top_k=40, min_p=0.05, typical_p=1.0, stream=False, stop=[], seed=None, response_format=None, max_tokens=None, presence_penalty=0.0, frequency_penalty=0.0, repeat_penalty=1.1, tfs_z=1.0, mirostat_mode=0, mirostat_tau=5.0, mirostat_eta=0.1, model=None, logits_processor=None, grammar=None, logit_bias=None)

Generate a chat completion from a list of messages.

Parameters:

  • messages (List[ChatCompletionRequestMessage]) –

    A list of messages to generate a response for.

  • functions (Optional[List[ChatCompletionFunction]], default: None ) –

    A list of functions to use for the chat completion.

  • function_call (Optional[ChatCompletionRequestFunctionCall], default: None ) –

    A function call to use for the chat completion.

  • tools (Optional[List[ChatCompletionTool]], default: None ) –

    A list of tools to use for the chat completion.

  • tool_choice (Optional[ChatCompletionToolChoiceOption], default: None ) –

    A tool choice to use for the chat completion.

  • temperature (float, default: 0.2 ) –

    The temperature to use for sampling.

  • top_p (float, default: 0.95 ) –

    The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751

  • top_k (int, default: 40 ) –

    The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751

  • min_p (float, default: 0.05 ) –

    The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841

  • typical_p (float, default: 1.0 ) –

    The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.

  • stream (bool, default: False ) –

    Whether to stream the results.

  • stop (Optional[Union[str, List[str]]], default: [] ) –

    A list of strings to stop generation when encountered.

  • seed (Optional[int], default: None ) –

    The seed to use for sampling.

  • response_format (Optional[ChatCompletionRequestResponseFormat], default: None ) –

    The response format to use for the chat completion. Use { "type": "json_object" } to contstrain output to only valid json.

  • max_tokens (Optional[int], default: None ) –

    The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.

  • presence_penalty (float, default: 0.0 ) –

    The penalty to apply to tokens based on their presence in the prompt.

  • frequency_penalty (float, default: 0.0 ) –

    The penalty to apply to tokens based on their frequency in the prompt.

  • repeat_penalty (float, default: 1.1 ) –

    The penalty to apply to repeated tokens.

  • tfs_z (float, default: 1.0 ) –

    The tail-free sampling parameter.

  • mirostat_mode (int, default: 0 ) –

    The mirostat sampling mode.

  • mirostat_tau (float, default: 5.0 ) –

    The mirostat sampling tau parameter.

  • mirostat_eta (float, default: 0.1 ) –

    The mirostat sampling eta parameter.

  • model (Optional[str], default: None ) –

    The name to use for the model in the completion object.

  • logits_processor (Optional[LogitsProcessorList], default: None ) –

    A list of logits processors to use.

  • grammar (Optional[LlamaGrammar], default: None ) –

    A grammar to use.

  • logit_bias (Optional[Dict[str, float]], default: None ) –

    A logit bias to use.

Returns:

Source code in llama_cpp/llama.py
def create_chat_completion(
    self,
    messages: List[ChatCompletionRequestMessage],
    functions: Optional[List[ChatCompletionFunction]] = None,
    function_call: Optional[ChatCompletionRequestFunctionCall] = None,
    tools: Optional[List[ChatCompletionTool]] = None,
    tool_choice: Optional[ChatCompletionToolChoiceOption] = None,
    temperature: float = 0.2,
    top_p: float = 0.95,
    top_k: int = 40,
    min_p: float = 0.05,
    typical_p: float = 1.0,
    stream: bool = False,
    stop: Optional[Union[str, List[str]]] = [],
    seed: Optional[int] = None,
    response_format: Optional[ChatCompletionRequestResponseFormat] = None,
    max_tokens: Optional[int] = None,
    presence_penalty: float = 0.0,
    frequency_penalty: float = 0.0,
    repeat_penalty: float = 1.1,
    tfs_z: float = 1.0,
    mirostat_mode: int = 0,
    mirostat_tau: float = 5.0,
    mirostat_eta: float = 0.1,
    model: Optional[str] = None,
    logits_processor: Optional[LogitsProcessorList] = None,
    grammar: Optional[LlamaGrammar] = None,
    logit_bias: Optional[Dict[str, float]] = None,
) -> Union[
    CreateChatCompletionResponse, Iterator[CreateChatCompletionStreamResponse]
]:
    """Generate a chat completion from a list of messages.

    Args:
        messages: A list of messages to generate a response for.
        functions: A list of functions to use for the chat completion.
        function_call: A function call to use for the chat completion.
        tools: A list of tools to use for the chat completion.
        tool_choice: A tool choice to use for the chat completion.
        temperature: The temperature to use for sampling.
        top_p: The top-p value to use for nucleus sampling. Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
        top_k: The top-k value to use for sampling. Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
        min_p: The min-p value to use for minimum p sampling. Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841
        typical_p: The typical-p value to use for sampling. Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
        stream: Whether to stream the results.
        stop: A list of strings to stop generation when encountered.
        seed: The seed to use for sampling.
        response_format: The response format to use for the chat completion. Use { "type": "json_object" } to contstrain output to only valid json.
        max_tokens: The maximum number of tokens to generate. If max_tokens <= 0 or None, the maximum number of tokens to generate is unlimited and depends on n_ctx.
        presence_penalty: The penalty to apply to tokens based on their presence in the prompt.
        frequency_penalty: The penalty to apply to tokens based on their frequency in the prompt.
        repeat_penalty: The penalty to apply to repeated tokens.
        tfs_z: The tail-free sampling parameter.
        mirostat_mode: The mirostat sampling mode.
        mirostat_tau: The mirostat sampling tau parameter.
        mirostat_eta: The mirostat sampling eta parameter.
        model: The name to use for the model in the completion object.
        logits_processor: A list of logits processors to use.
        grammar: A grammar to use.
        logit_bias: A logit bias to use.

    Returns:
        Generated chat completion or a stream of chat completion chunks.
    """
    handler = self.chat_handler or llama_chat_format.get_chat_completion_handler(
        self.chat_format
    )
    return handler(
        llama=self,
        messages=messages,
        functions=functions,
        function_call=function_call,
        tools=tools,
        tool_choice=tool_choice,
        temperature=temperature,
        top_p=top_p,
        top_k=top_k,
        min_p=min_p,
        typical_p=typical_p,
        stream=stream,
        stop=stop,
        seed=seed,
        response_format=response_format,
        max_tokens=max_tokens,
        presence_penalty=presence_penalty,
        frequency_penalty=frequency_penalty,
        repeat_penalty=repeat_penalty,
        tfs_z=tfs_z,
        mirostat_mode=mirostat_mode,
        mirostat_tau=mirostat_tau,
        mirostat_eta=mirostat_eta,
        model=model,
        logits_processor=logits_processor,
        grammar=grammar,
        logit_bias=logit_bias,
    )

set_cache(cache)

Set the cache.

Parameters:

  • cache (Optional[BaseLlamaCache]) –

    The cache to set.

Source code in llama_cpp/llama.py
def set_cache(self, cache: Optional[BaseLlamaCache]):
    """Set the cache.

    Args:
        cache: The cache to set.
    """
    self.cache = cache

save_state()

Source code in llama_cpp/llama.py
def save_state(self) -> LlamaState:
    assert self._ctx.ctx is not None
    if self.verbose:
        print("Llama.save_state: saving llama state", file=sys.stderr)
    state_size = llama_cpp.llama_get_state_size(self._ctx.ctx)
    if self.verbose:
        print(f"Llama.save_state: got state size: {state_size}", file=sys.stderr)
    llama_state = (llama_cpp.c_uint8 * int(state_size))()
    if self.verbose:
        print("Llama.save_state: allocated state", file=sys.stderr)
    n_bytes = llama_cpp.llama_copy_state_data(self._ctx.ctx, llama_state)
    if self.verbose:
        print(f"Llama.save_state: copied llama state: {n_bytes}", file=sys.stderr)
    if int(n_bytes) > int(state_size):
        raise RuntimeError("Failed to copy llama state data")
    llama_state_compact = (llama_cpp.c_uint8 * int(n_bytes))()
    llama_cpp.ctypes.memmove(llama_state_compact, llama_state, int(n_bytes))
    if self.verbose:
        print(
            f"Llama.save_state: saving {n_bytes} bytes of llama state",
            file=sys.stderr,
        )
    return LlamaState(
        scores=self.scores.copy(),
        input_ids=self.input_ids.copy(),
        n_tokens=self.n_tokens,
        llama_state=bytes(llama_state_compact),
        llama_state_size=n_bytes,
    )

load_state(state)

Source code in llama_cpp/llama.py
def load_state(self, state: LlamaState) -> None:
    assert self._ctx.ctx is not None
    self.scores = state.scores.copy()
    self.input_ids = state.input_ids.copy()
    self.n_tokens = state.n_tokens
    state_size = state.llama_state_size
    LLamaStateArrayType = llama_cpp.c_uint8 * state_size
    llama_state = LLamaStateArrayType.from_buffer_copy(state.llama_state)

    if llama_cpp.llama_set_state_data(self._ctx.ctx, llama_state) != state_size:
        raise RuntimeError("Failed to set llama state data")

token_bos()

Return the beginning-of-sequence token.

Source code in llama_cpp/llama.py
def token_bos(self) -> int:
    """Return the beginning-of-sequence token."""
    return self._model.token_bos()

token_eos()

Return the end-of-sequence token.

Source code in llama_cpp/llama.py
def token_eos(self) -> int:
    """Return the end-of-sequence token."""
    return self._model.token_eos()

llama_cpp.LlamaGrammar

Keeps reference counts of all the arguments, so that they are not garbage collected by Python.

Source code in llama_cpp/llama_grammar.py