llama.cpp/llama_cpp/llama_cpp.py

943 lines
32 KiB
Python
Raw Normal View History

import sys
import os
2023-03-23 09:33:06 +00:00
import ctypes
2023-04-11 15:59:03 +00:00
from ctypes import (
c_int,
c_float,
c_char_p,
c_void_p,
c_bool,
POINTER,
2023-05-05 17:54:22 +00:00
_Pointer, # type: ignore
2023-04-11 15:59:03 +00:00
Structure,
Array,
c_uint8,
c_size_t,
)
2023-03-23 09:33:06 +00:00
import pathlib
from typing import List
2023-03-23 09:33:06 +00:00
2023-04-11 15:59:03 +00:00
2023-03-23 09:33:06 +00:00
# Load the library
def _load_shared_library(lib_base_name: str):
# Construct the paths to the possible shared library names
_base_path = pathlib.Path(__file__).parent.resolve()
# Searching for the library in the current directory under the name "libllama" (default name
# for llamacpp) and "llama" (default name for this repo)
_lib_paths: List[pathlib.Path] = []
# Determine the file extension based on the platform
if sys.platform.startswith("linux"):
_lib_paths += [
_base_path / f"lib{lib_base_name}.so",
]
elif sys.platform == "darwin":
_lib_paths += [
_base_path / f"lib{lib_base_name}.so",
_base_path / f"lib{lib_base_name}.dylib",
]
elif sys.platform == "win32":
_lib_paths += [
_base_path / f"{lib_base_name}.dll",
]
else:
raise RuntimeError("Unsupported platform")
2023-04-11 15:59:03 +00:00
if "LLAMA_CPP_LIB" in os.environ:
lib_base_name = os.environ["LLAMA_CPP_LIB"]
2023-04-10 15:12:25 +00:00
_lib = pathlib.Path(lib_base_name)
_base_path = _lib.parent.resolve()
_lib_paths = [_lib.resolve()]
cdll_args = dict() # type: ignore
# Add the library directory to the DLL search path on Windows (if needed)
if sys.platform == "win32" and sys.version_info >= (3, 8):
os.add_dll_directory(str(_base_path))
2023-05-17 19:26:38 +00:00
if "CUDA_PATH" in os.environ:
os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "bin"))
os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "lib"))
cdll_args["winmode"] = 0
# Try to load the shared library, handling potential errors
for _lib_path in _lib_paths:
if _lib_path.exists():
try:
return ctypes.CDLL(str(_lib_path), **cdll_args)
except Exception as e:
raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}")
2023-04-11 15:59:03 +00:00
raise FileNotFoundError(
f"Shared library with base name '{lib_base_name}' not found"
)
# Specify the base name of the shared library to load
_lib_base_name = "llama"
# Load the library
_lib = _load_shared_library(_lib_base_name)
2023-03-23 09:33:06 +00:00
2023-05-21 21:47:21 +00:00
# Misc
c_float_p = POINTER(c_float)
c_uint8_p = POINTER(c_uint8)
c_size_t_p = POINTER(c_size_t)
# llama.h bindings
2023-06-06 20:23:55 +00:00
GGML_USE_CUBLAS = hasattr(_lib, "ggml_init_cublas")
GGML_CUDA_MAX_DEVICES = ctypes.c_int(16)
LLAMA_MAX_DEVICES = GGML_CUDA_MAX_DEVICES if GGML_USE_CUBLAS else ctypes.c_int(1)
2023-05-21 21:47:21 +00:00
# #define LLAMA_FILE_MAGIC_GGJT 0x67676a74u // 'ggjt'
LLAMA_FILE_MAGIC_GGJT = ctypes.c_uint(0x67676A74)
# #define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
LLAMA_FILE_MAGIC_GGLA = ctypes.c_uint(0x67676C61)
# #define LLAMA_FILE_MAGIC_GGMF 0x67676d66u // 'ggmf'
LLAMA_FILE_MAGIC_GGMF = ctypes.c_uint(0x67676D66)
# #define LLAMA_FILE_MAGIC_GGML 0x67676d6cu // 'ggml'
LLAMA_FILE_MAGIC_GGML = ctypes.c_uint(0x67676D6C)
# #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
LLAMA_FILE_MAGIC_GGSN = ctypes.c_uint(0x6767736E)
# #define LLAMA_FILE_VERSION 3
LLAMA_FILE_VERSION = c_int(3)
LLAMA_FILE_MAGIC = LLAMA_FILE_MAGIC_GGJT
LLAMA_FILE_MAGIC_UNVERSIONED = LLAMA_FILE_MAGIC_GGML
LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN
2023-05-05 18:05:31 +00:00
LLAMA_SESSION_VERSION = c_int(1)
2023-05-01 14:44:28 +00:00
2023-05-21 21:47:21 +00:00
# struct llama_context;
llama_context_p = c_void_p
2023-05-21 21:47:21 +00:00
# typedef int llama_token;
2023-03-23 09:33:06 +00:00
llama_token = c_int
llama_token_p = POINTER(llama_token)
2023-03-24 18:35:41 +00:00
2023-05-21 21:47:21 +00:00
# typedef struct llama_token_data {
# llama_token id; // token id
# float logit; // log-odds of the token
# float p; // probability of the token
# } llama_token_data;
2023-03-23 09:33:06 +00:00
class llama_token_data(Structure):
_fields_ = [
2023-05-21 21:47:21 +00:00
("id", llama_token),
("logit", c_float),
("p", c_float),
2023-03-23 09:33:06 +00:00
]
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
llama_token_data_p = POINTER(llama_token_data)
2023-05-01 14:44:28 +00:00
2023-05-21 21:47:21 +00:00
# typedef struct llama_token_data_array {
# llama_token_data * data;
# size_t size;
# bool sorted;
# } llama_token_data_array;
2023-05-01 14:44:28 +00:00
class llama_token_data_array(Structure):
_fields_ = [
("data", llama_token_data_p),
("size", c_size_t),
("sorted", c_bool),
]
llama_token_data_array_p = POINTER(llama_token_data_array)
2023-05-21 21:47:21 +00:00
# typedef void (*llama_progress_callback)(float progress, void *ctx);
2023-03-29 01:10:23 +00:00
llama_progress_callback = ctypes.CFUNCTYPE(None, c_float, c_void_p)
2023-03-24 18:35:41 +00:00
2023-03-25 20:26:03 +00:00
2023-05-21 21:47:21 +00:00
# struct llama_context_params {
2023-06-06 20:23:55 +00:00
# int n_ctx; // text context
# int n_batch; // prompt processing batch size
# int n_gpu_layers; // number of layers to store in VRAM
# int main_gpu; // the GPU that is used for scratch and small tensors
# float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
2023-06-15 01:40:13 +00:00
# bool low_vram; // if true, reduce VRAM usage at the cost of performance
2023-06-06 20:23:55 +00:00
# int seed; // RNG seed, -1 for random
2023-05-21 21:47:21 +00:00
# bool f16_kv; // use fp16 for KV cache
# bool logits_all; // the llama_eval() call computes all logits, not just the last one
# bool vocab_only; // only load the vocabulary, no weights
# bool use_mmap; // use mmap if possible
# bool use_mlock; // force system to keep model in RAM
# bool embedding; // embedding mode only
2023-05-21 21:47:21 +00:00
# // called with a progress value between 0 and 1, pass NULL to disable
# llama_progress_callback progress_callback;
# // context pointer passed to the progress callback
# void * progress_callback_user_data;
# };
2023-03-23 09:33:06 +00:00
class llama_context_params(Structure):
_fields_ = [
2023-05-21 21:47:21 +00:00
("n_ctx", c_int),
2023-06-06 20:23:55 +00:00
("n_batch", c_int),
2023-05-21 21:47:21 +00:00
("n_gpu_layers", c_int),
2023-06-06 20:23:55 +00:00
("main_gpu", c_int),
("tensor_split", c_float * LLAMA_MAX_DEVICES.value),
2023-06-15 01:40:13 +00:00
("low_vram", c_bool),
2023-05-21 21:47:21 +00:00
("seed", c_int),
("f16_kv", c_bool),
2023-03-24 18:35:41 +00:00
(
"logits_all",
c_bool,
2023-05-21 21:47:21 +00:00
),
("vocab_only", c_bool),
("use_mmap", c_bool),
("use_mlock", c_bool),
("embedding", c_bool),
2023-03-25 16:12:09 +00:00
("progress_callback", llama_progress_callback),
("progress_callback_user_data", c_void_p),
2023-03-23 09:33:06 +00:00
]
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
llama_context_params_p = POINTER(llama_context_params)
2023-05-21 21:47:21 +00:00
# enum llama_ftype {
# LLAMA_FTYPE_ALL_F32 = 0,
# LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
# // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed
# // LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed
# LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors
2023-06-05 22:17:29 +00:00
# LLAMA_FTYPE_MOSTLY_Q2_K = 10,// except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q3_K_S = 11,// except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q3_K_M = 12,// except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q3_K_L = 13,// except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_K_S = 14,// except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_K_M = 15,// except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q5_K_S = 16,// except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q5_K_M = 17,// except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q6_K = 18,// except 1d tensors
2023-05-21 21:47:21 +00:00
# };
2023-05-05 18:05:31 +00:00
LLAMA_FTYPE_ALL_F32 = c_int(0)
2023-05-21 21:47:21 +00:00
LLAMA_FTYPE_MOSTLY_F16 = c_int(1)
LLAMA_FTYPE_MOSTLY_Q4_0 = c_int(2)
LLAMA_FTYPE_MOSTLY_Q4_1 = c_int(3)
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = c_int(4)
LLAMA_FTYPE_MOSTLY_Q8_0 = c_int(7)
LLAMA_FTYPE_MOSTLY_Q5_0 = c_int(8)
LLAMA_FTYPE_MOSTLY_Q5_1 = c_int(9)
2023-06-05 22:17:29 +00:00
LLAMA_FTYPE_MOSTLY_Q2_K = c_int(10)
LLAMA_FTYPE_MOSTLY_Q3_K_S = c_int(11)
LLAMA_FTYPE_MOSTLY_Q3_K_M = c_int(12)
LLAMA_FTYPE_MOSTLY_Q3_K_L = c_int(13)
LLAMA_FTYPE_MOSTLY_Q4_K_S = c_int(14)
LLAMA_FTYPE_MOSTLY_Q4_K_M = c_int(15)
LLAMA_FTYPE_MOSTLY_Q5_K_S = c_int(16)
LLAMA_FTYPE_MOSTLY_Q5_K_M = c_int(17)
LLAMA_FTYPE_MOSTLY_Q6_K = c_int(18)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
2023-06-10 16:17:38 +00:00
# // model quantization parameters
# typedef struct llama_model_quantize_params {
# int nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
# enum llama_ftype ftype; // quantize to this llama_ftype
# bool allow_requantize; // allow quantizing non-f32/f16 tensors
# bool quantize_output_tensor; // quantize output.weight
# } llama_model_quantize_params;
class llama_model_quantize_params(Structure):
_fields_ = [
("nthread", c_int),
("ftype", c_int),
("allow_requantize", c_bool),
("quantize_output_tensor", c_bool),
]
2023-05-21 21:47:21 +00:00
# LLAMA_API struct llama_context_params llama_context_default_params();
2023-03-23 09:33:06 +00:00
def llama_context_default_params() -> llama_context_params:
2023-03-29 01:10:23 +00:00
return _lib.llama_context_default_params()
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_context_default_params.argtypes = []
_lib.llama_context_default_params.restype = llama_context_params
2023-04-11 15:59:03 +00:00
2023-06-10 16:17:38 +00:00
# LLAMA_API struct llama_model_quantize_params llama_model_quantize_default_params();
def llama_model_quantize_default_params() -> llama_model_quantize_params:
return _lib.llama_model_quantize_default_params()
_lib.llama_model_quantize_default_params.argtypes = []
_lib.llama_model_quantize_default_params.restype = llama_model_quantize_params
2023-05-21 21:47:21 +00:00
# LLAMA_API bool llama_mmap_supported();
def llama_mmap_supported() -> bool:
2023-04-10 02:01:33 +00:00
return _lib.llama_mmap_supported()
2023-04-11 15:59:03 +00:00
2023-04-10 02:01:33 +00:00
_lib.llama_mmap_supported.argtypes = []
_lib.llama_mmap_supported.restype = c_bool
2023-04-11 15:59:03 +00:00
2023-05-21 21:47:21 +00:00
# LLAMA_API bool llama_mlock_supported();
def llama_mlock_supported() -> bool:
2023-04-10 02:01:33 +00:00
return _lib.llama_mlock_supported()
2023-04-11 15:59:03 +00:00
2023-04-10 02:01:33 +00:00
_lib.llama_mlock_supported.argtypes = []
_lib.llama_mlock_supported.restype = c_bool
2023-03-24 18:59:29 +00:00
2023-04-11 15:59:03 +00:00
2023-05-21 21:47:21 +00:00
# // TODO: not great API - very likely to change
# // Initialize the llama + ggml backend
# // Call once at the start of the program
# LLAMA_API void llama_init_backend();
def llama_init_backend():
return _lib.llama_init_backend()
_lib.llama_init_backend.argtypes = []
_lib.llama_init_backend.restype = None
# LLAMA_API int64_t llama_time_us();
def llama_time_us() -> int:
return _lib.llama_time_us()
_lib.llama_time_us.argtypes = []
_lib.llama_time_us.restype = ctypes.c_int64
# // Various functions for loading a ggml llama model.
# // Allocate (almost) all memory needed for the model.
# // Return NULL on failure
# LLAMA_API struct llama_context * llama_init_from_file(
# const char * path_model,
# struct llama_context_params params);
2023-03-24 18:35:41 +00:00
def llama_init_from_file(
path_model: bytes, params: llama_context_params
) -> llama_context_p:
return _lib.llama_init_from_file(path_model, params)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_init_from_file.argtypes = [c_char_p, llama_context_params]
_lib.llama_init_from_file.restype = llama_context_p
2023-03-24 18:59:29 +00:00
# Frees all allocated memory
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_free(struct llama_context * ctx);
2023-03-23 09:33:06 +00:00
def llama_free(ctx: llama_context_p):
return _lib.llama_free(ctx)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_free.argtypes = [llama_context_p]
_lib.llama_free.restype = None
2023-03-24 18:59:29 +00:00
2023-06-10 16:17:38 +00:00
# // Returns 0 on success
2023-05-21 21:47:21 +00:00
# LLAMA_API int llama_model_quantize(
# const char * fname_inp,
# const char * fname_out,
2023-06-10 16:17:38 +00:00
# const llama_model_quantize_params * params);
2023-04-22 23:50:28 +00:00
def llama_model_quantize(
2023-06-10 16:17:38 +00:00
fname_inp: bytes,
fname_out: bytes,
params, # type: POINTER(llama_model_quantize_params) # type: ignore
) -> int:
2023-06-10 16:17:38 +00:00
return _lib.llama_model_quantize(fname_inp, fname_out, params)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
2023-06-10 16:17:38 +00:00
_lib.llama_model_quantize.argtypes = [
c_char_p,
c_char_p,
POINTER(llama_model_quantize_params),
]
_lib.llama_model_quantize.restype = c_int
2023-04-11 15:59:03 +00:00
# Apply a LoRA adapter to a loaded model
# path_base_model is the path to a higher quality model to use as a base for
# the layers modified by the adapter. Can be NULL to use the current loaded model.
# The model needs to be reloaded before applying a new adapter, otherwise the adapter
# will be applied on top of the previous one
# Returns 0 on success
2023-05-21 21:47:21 +00:00
# LLAMA_API int llama_apply_lora_from_file(
# struct llama_context * ctx,
# const char * path_lora,
# const char * path_base_model,
# int n_threads);
def llama_apply_lora_from_file(
2023-04-22 23:50:28 +00:00
ctx: llama_context_p,
2023-05-05 18:05:31 +00:00
path_lora: c_char_p,
path_base_model: c_char_p,
2023-04-22 23:50:28 +00:00
n_threads: c_int,
) -> int:
return _lib.llama_apply_lora_from_file(ctx, path_lora, path_base_model, n_threads)
_lib.llama_apply_lora_from_file.argtypes = [llama_context_p, c_char_p, c_char_p, c_int]
_lib.llama_apply_lora_from_file.restype = c_int
# Returns the number of tokens in the KV cache
2023-05-21 21:47:21 +00:00
# LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
def llama_get_kv_cache_token_count(ctx: llama_context_p) -> int:
return _lib.llama_get_kv_cache_token_count(ctx)
2023-04-11 15:59:03 +00:00
_lib.llama_get_kv_cache_token_count.argtypes = [llama_context_p]
_lib.llama_get_kv_cache_token_count.restype = c_int
2023-04-28 19:32:43 +00:00
2023-04-27 00:00:54 +00:00
# Sets the current rng seed.
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed);
2023-04-27 00:00:54 +00:00
def llama_set_rng_seed(ctx: llama_context_p, seed: c_int):
return _lib.llama_set_rng_seed(ctx, seed)
2023-04-28 19:32:43 +00:00
2023-04-27 00:00:54 +00:00
_lib.llama_set_rng_seed.argtypes = [llama_context_p, c_int]
_lib.llama_set_rng_seed.restype = None
2023-04-28 19:32:43 +00:00
2023-05-03 13:33:30 +00:00
# Returns the maximum size in bytes of the state (rng, logits, embedding
# and kv_cache) - will often be smaller after compacting tokens
2023-05-21 21:47:21 +00:00
# LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);
def llama_get_state_size(ctx: llama_context_p) -> int:
2023-04-22 23:50:28 +00:00
return _lib.llama_get_state_size(ctx)
_lib.llama_get_state_size.argtypes = [llama_context_p]
_lib.llama_get_state_size.restype = c_size_t
# Copies the state to the specified destination address.
# Destination needs to have allocated enough memory.
# Returns the number of bytes copied
2023-05-21 21:47:21 +00:00
# LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
2023-05-05 18:12:26 +00:00
def llama_copy_state_data(
2023-05-14 04:04:22 +00:00
ctx: llama_context_p, dst # type: Array[c_uint8]
2023-05-07 23:30:14 +00:00
) -> int:
2023-05-14 04:04:22 +00:00
return _lib.llama_copy_state_data(ctx, dst)
2023-04-22 23:50:28 +00:00
2023-05-05 18:04:12 +00:00
_lib.llama_copy_state_data.argtypes = [llama_context_p, c_uint8_p]
2023-04-22 23:50:28 +00:00
_lib.llama_copy_state_data.restype = c_size_t
# Set the state reading from the specified address
# Returns the number of bytes read
2023-05-21 21:47:21 +00:00
# LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
def llama_set_state_data(
ctx: llama_context_p, src # type: Array[c_uint8]
2023-05-07 23:30:14 +00:00
) -> int:
2023-04-22 23:50:28 +00:00
return _lib.llama_set_state_data(ctx, src)
2023-05-05 18:04:12 +00:00
_lib.llama_set_state_data.argtypes = [llama_context_p, c_uint8_p]
2023-04-22 23:50:28 +00:00
_lib.llama_set_state_data.restype = c_size_t
2023-04-28 19:32:43 +00:00
# Save/load session file
2023-05-21 21:47:21 +00:00
# LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
2023-04-28 19:32:43 +00:00
def llama_load_session_file(
ctx: llama_context_p,
path_session: bytes,
2023-05-05 18:12:26 +00:00
tokens_out, # type: Array[llama_token]
2023-04-28 19:32:43 +00:00
n_token_capacity: c_size_t,
2023-05-05 18:12:26 +00:00
n_token_count_out, # type: _Pointer[c_size_t]
) -> int:
2023-04-28 19:32:43 +00:00
return _lib.llama_load_session_file(
ctx, path_session, tokens_out, n_token_capacity, n_token_count_out
)
_lib.llama_load_session_file.argtypes = [
llama_context_p,
c_char_p,
llama_token_p,
c_size_t,
2023-05-05 18:04:12 +00:00
c_size_t_p,
2023-04-28 19:32:43 +00:00
]
_lib.llama_load_session_file.restype = c_size_t
2023-05-21 21:47:21 +00:00
# LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
2023-04-28 19:32:43 +00:00
def llama_save_session_file(
ctx: llama_context_p,
path_session: bytes,
2023-05-05 18:12:26 +00:00
tokens, # type: Array[llama_token]
n_token_count: c_size_t,
) -> int:
2023-04-28 19:32:43 +00:00
return _lib.llama_save_session_file(ctx, path_session, tokens, n_token_count)
_lib.llama_save_session_file.argtypes = [
llama_context_p,
c_char_p,
llama_token_p,
c_size_t,
]
_lib.llama_save_session_file.restype = c_size_t
# Run the llama inference to obtain the logits and probabilities for the next token.
# tokens + n_tokens is the provided batch of new tokens to process
# n_past is the number of tokens to use from previous eval calls
# Returns 0 on success
2023-05-21 21:47:21 +00:00
# LLAMA_API int llama_eval(
# struct llama_context * ctx,
# const llama_token * tokens,
# int n_tokens,
# int n_past,
# int n_threads);
2023-03-24 18:35:41 +00:00
def llama_eval(
ctx: llama_context_p,
2023-05-05 18:12:26 +00:00
tokens, # type: Array[llama_token]
2023-03-24 18:35:41 +00:00
n_tokens: c_int,
n_past: c_int,
n_threads: c_int,
) -> int:
return _lib.llama_eval(ctx, tokens, n_tokens, n_past, n_threads)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_eval.argtypes = [llama_context_p, llama_token_p, c_int, c_int, c_int]
_lib.llama_eval.restype = c_int
# Convert the provided text into tokens.
# The tokens pointer must be large enough to hold the resulting tokens.
# Returns the number of tokens on success, no more than n_max_tokens
# Returns a negative number on failure - the number of tokens that would have been returned
# TODO: not sure if correct
2023-05-21 21:47:21 +00:00
# LLAMA_API int llama_tokenize(
# struct llama_context * ctx,
# const char * text,
# llama_token * tokens,
# int n_max_tokens,
# bool add_bos);
2023-03-24 18:35:41 +00:00
def llama_tokenize(
ctx: llama_context_p,
text: bytes,
2023-05-05 18:12:26 +00:00
tokens, # type: Array[llama_token]
2023-03-24 18:35:41 +00:00
n_max_tokens: c_int,
add_bos: c_bool,
) -> int:
return _lib.llama_tokenize(ctx, text, tokens, n_max_tokens, add_bos)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_tokenize.argtypes = [llama_context_p, c_char_p, llama_token_p, c_int, c_bool]
_lib.llama_tokenize.restype = c_int
2023-05-21 21:47:21 +00:00
# LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
def llama_n_vocab(ctx: llama_context_p) -> int:
return _lib.llama_n_vocab(ctx)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_n_vocab.argtypes = [llama_context_p]
_lib.llama_n_vocab.restype = c_int
2023-05-21 21:47:21 +00:00
# LLAMA_API int llama_n_ctx (const struct llama_context * ctx);
def llama_n_ctx(ctx: llama_context_p) -> int:
return _lib.llama_n_ctx(ctx)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_n_ctx.argtypes = [llama_context_p]
_lib.llama_n_ctx.restype = c_int
2023-03-24 18:59:29 +00:00
2023-05-21 21:47:21 +00:00
# LLAMA_API int llama_n_embd (const struct llama_context * ctx);
def llama_n_embd(ctx: llama_context_p) -> int:
return _lib.llama_n_embd(ctx)
2023-03-25 20:26:03 +00:00
_lib.llama_n_embd.argtypes = [llama_context_p]
_lib.llama_n_embd.restype = c_int
2023-06-15 01:40:13 +00:00
# // Get the vocabulary as output parameters.
# // Returns number of results.
# LLAMA_API int llama_get_vocab(
# const struct llama_context * ctx,
# const char * * strings,
# float * scores,
# int capacity);
def llama_get_vocab(
ctx: llama_context_p,
strings, # type: Array[c_char_p] # type: ignore
scores, # type: Array[c_float] # type: ignore
capacity: c_int,
) -> int:
return _lib.llama_get_vocab(ctx, strings, scores, capacity)
_lib.llama_get_vocab.argtypes = [llama_context_p, c_char_p, c_float, c_int]
_lib.llama_get_vocab.restype = c_int
# Token logits obtained from the last call to llama_eval()
# The logits for the last token are stored in the last row
# Can be mutated in order to change the probabilities of the next token
# Rows: n_tokens
# Cols: n_vocab
2023-05-21 21:47:21 +00:00
# LLAMA_API float * llama_get_logits(struct llama_context * ctx);
def llama_get_logits(
ctx: llama_context_p,
): # type: (...) -> Array[float] # type: ignore
return _lib.llama_get_logits(ctx)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_get_logits.argtypes = [llama_context_p]
2023-05-05 18:04:12 +00:00
_lib.llama_get_logits.restype = c_float_p
2023-03-24 18:59:29 +00:00
# Get the embeddings for the input
# shape: [n_embd] (1-dimensional)
2023-05-21 21:47:21 +00:00
# LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
def llama_get_embeddings(
ctx: llama_context_p,
): # type: (...) -> Array[float] # type: ignore
return _lib.llama_get_embeddings(ctx)
2023-03-24 18:59:29 +00:00
_lib.llama_get_embeddings.argtypes = [llama_context_p]
2023-05-05 18:04:12 +00:00
_lib.llama_get_embeddings.restype = c_float_p
2023-03-24 18:59:29 +00:00
# Token Id -> String. Uses the vocabulary in the provided context
2023-05-21 21:47:21 +00:00
# LLAMA_API const char * llama_token_to_str(const struct llama_context * ctx, llama_token token);
2023-03-31 07:25:12 +00:00
def llama_token_to_str(ctx: llama_context_p, token: llama_token) -> bytes:
return _lib.llama_token_to_str(ctx, token)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_token_to_str.argtypes = [llama_context_p, llama_token]
_lib.llama_token_to_str.restype = c_char_p
# Special tokens
2023-03-24 18:59:29 +00:00
2023-05-21 21:47:21 +00:00
# LLAMA_API llama_token llama_token_bos();
def llama_token_bos() -> int:
return _lib.llama_token_bos()
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_token_bos.argtypes = []
_lib.llama_token_bos.restype = llama_token
2023-05-21 21:47:21 +00:00
# LLAMA_API llama_token llama_token_eos();
def llama_token_eos() -> int:
return _lib.llama_token_eos()
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_token_eos.argtypes = []
_lib.llama_token_eos.restype = llama_token
2023-05-21 21:47:21 +00:00
# LLAMA_API llama_token llama_token_nl();
def llama_token_nl() -> int:
2023-05-01 14:44:28 +00:00
return _lib.llama_token_nl()
_lib.llama_token_nl.argtypes = []
_lib.llama_token_nl.restype = llama_token
# Sampling functions
2023-05-01 18:02:06 +00:00
# @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
2023-05-01 14:44:28 +00:00
def llama_sample_repetition_penalty(
ctx: llama_context_p,
2023-05-05 18:12:26 +00:00
candidates, # type: _Pointer[llama_token_data_array]
last_tokens_data, # type: Array[llama_token]
2023-05-01 14:44:28 +00:00
last_tokens_size: c_int,
penalty: c_float,
2023-05-01 18:02:06 +00:00
):
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_repetition_penalty(
ctx, candidates, last_tokens_data, last_tokens_size, penalty
)
_lib.llama_sample_repetition_penalty.argtypes = [
llama_context_p,
llama_token_data_array_p,
llama_token_p,
c_int,
c_float,
]
2023-05-01 18:02:06 +00:00
_lib.llama_sample_repetition_penalty.restype = None
2023-05-01 14:44:28 +00:00
2023-05-01 18:02:06 +00:00
# @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
2023-05-01 14:44:28 +00:00
def llama_sample_frequency_and_presence_penalties(
2023-03-24 18:35:41 +00:00
ctx: llama_context_p,
2023-05-05 18:12:26 +00:00
candidates, # type: _Pointer[llama_token_data_array]
last_tokens_data, # type: Array[llama_token]
2023-05-01 14:44:28 +00:00
last_tokens_size: c_int,
alpha_frequency: c_float,
alpha_presence: c_float,
2023-05-01 18:02:06 +00:00
):
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_frequency_and_presence_penalties(
ctx,
candidates,
last_tokens_data,
last_tokens_size,
alpha_frequency,
alpha_presence,
2023-03-24 18:35:41 +00:00
)
2023-03-23 09:33:06 +00:00
2023-05-01 14:44:28 +00:00
_lib.llama_sample_frequency_and_presence_penalties.argtypes = [
llama_context_p,
2023-05-01 14:44:28 +00:00
llama_token_data_array_p,
llama_token_p,
c_int,
2023-05-01 14:44:28 +00:00
c_float,
c_float,
]
2023-05-01 18:02:06 +00:00
_lib.llama_sample_frequency_and_presence_penalties.restype = None
2023-05-01 14:44:28 +00:00
2023-05-01 18:02:06 +00:00
# @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
2023-05-05 18:12:26 +00:00
def llama_sample_softmax(
ctx: llama_context_p, candidates # type: _Pointer[llama_token_data]
):
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_softmax(ctx, candidates)
_lib.llama_sample_softmax.argtypes = [
llama_context_p,
llama_token_data_array_p,
]
2023-05-01 18:02:06 +00:00
_lib.llama_sample_softmax.restype = None
2023-05-01 14:44:28 +00:00
2023-05-01 18:02:06 +00:00
# @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
2023-05-01 18:47:55 +00:00
def llama_sample_top_k(
2023-05-05 17:54:22 +00:00
ctx: llama_context_p,
2023-05-05 18:12:26 +00:00
candidates, # type: _Pointer[llama_token_data_array]
2023-05-05 17:54:22 +00:00
k: c_int,
2023-05-07 04:12:47 +00:00
min_keep: c_size_t,
2023-05-01 18:47:55 +00:00
):
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_top_k(ctx, candidates, k, min_keep)
_lib.llama_sample_top_k.argtypes = [
llama_context_p,
llama_token_data_array_p,
c_int,
2023-05-01 18:47:55 +00:00
c_size_t,
2023-05-01 14:44:28 +00:00
]
2023-05-01 18:02:06 +00:00
_lib.llama_sample_top_k.restype = None
2023-05-01 14:44:28 +00:00
2023-05-01 18:02:06 +00:00
# @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
2023-05-01 18:47:55 +00:00
def llama_sample_top_p(
2023-05-05 17:54:22 +00:00
ctx: llama_context_p,
2023-05-05 18:12:26 +00:00
candidates, # type: _Pointer[llama_token_data_array]
2023-05-05 17:54:22 +00:00
p: c_float,
2023-05-07 04:12:47 +00:00
min_keep: c_size_t,
2023-05-01 18:47:55 +00:00
):
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_top_p(ctx, candidates, p, min_keep)
_lib.llama_sample_top_p.argtypes = [
llama_context_p,
llama_token_data_array_p,
c_float,
2023-05-01 18:47:55 +00:00
c_size_t,
2023-05-01 14:44:28 +00:00
]
2023-05-01 18:02:06 +00:00
_lib.llama_sample_top_p.restype = None
2023-05-01 14:44:28 +00:00
2023-05-01 18:02:06 +00:00
# @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
2023-05-01 14:44:28 +00:00
def llama_sample_tail_free(
ctx: llama_context_p,
2023-05-05 18:12:26 +00:00
candidates, # type: _Pointer[llama_token_data_array]
z: c_float,
2023-05-07 04:12:47 +00:00
min_keep: c_size_t,
2023-05-01 18:02:06 +00:00
):
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_tail_free(ctx, candidates, z, min_keep)
_lib.llama_sample_tail_free.argtypes = [
llama_context_p,
llama_token_data_array_p,
c_float,
2023-05-01 18:47:55 +00:00
c_size_t,
2023-05-01 14:44:28 +00:00
]
2023-05-01 18:02:06 +00:00
_lib.llama_sample_tail_free.restype = None
2023-05-01 14:44:28 +00:00
2023-05-01 18:02:06 +00:00
# @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
2023-05-01 18:47:55 +00:00
def llama_sample_typical(
ctx: llama_context_p,
2023-05-05 18:12:26 +00:00
candidates, # type: _Pointer[llama_token_data_array]
2023-05-05 17:54:22 +00:00
p: c_float,
2023-05-07 04:12:47 +00:00
min_keep: c_size_t,
2023-05-01 18:47:55 +00:00
):
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_typical(ctx, candidates, p, min_keep)
_lib.llama_sample_typical.argtypes = [
llama_context_p,
llama_token_data_array_p,
c_float,
2023-05-01 18:47:55 +00:00
c_size_t,
2023-05-01 14:44:28 +00:00
]
2023-05-01 18:02:06 +00:00
_lib.llama_sample_typical.restype = None
2023-05-01 14:44:28 +00:00
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
def llama_sample_temperature(
2023-05-05 18:12:26 +00:00
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
temp: c_float,
):
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_temperature(ctx, candidates, temp)
_lib.llama_sample_temperature.argtypes = [
llama_context_p,
llama_token_data_array_p,
2023-03-29 01:10:23 +00:00
c_float,
2023-05-01 14:44:28 +00:00
]
2023-05-01 18:02:06 +00:00
_lib.llama_sample_temperature.restype = None
2023-05-01 14:44:28 +00:00
2023-05-01 18:02:06 +00:00
# @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
# @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
# @param 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.
# @param 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.
# @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
# @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
2023-05-21 21:47:21 +00:00
# LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
2023-05-01 14:44:28 +00:00
def llama_sample_token_mirostat(
ctx: llama_context_p,
2023-05-05 18:12:26 +00:00
candidates, # type: _Pointer[llama_token_data_array]
tau: c_float,
2023-05-05 17:54:22 +00:00
eta: c_float,
m: c_int,
2023-05-05 18:12:26 +00:00
mu, # type: _Pointer[c_float]
) -> int:
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_token_mirostat(ctx, candidates, tau, eta, m, mu)
_lib.llama_sample_token_mirostat.argtypes = [
llama_context_p,
llama_token_data_array_p,
c_float,
c_float,
c_int,
2023-05-05 18:04:12 +00:00
c_float_p,
2023-05-01 14:44:28 +00:00
]
_lib.llama_sample_token_mirostat.restype = llama_token
2023-05-01 18:02:06 +00:00
# @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
# @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
# @param 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.
# @param 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.
# @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
2023-05-21 21:47:21 +00:00
# LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
2023-05-01 14:44:28 +00:00
def llama_sample_token_mirostat_v2(
ctx: llama_context_p,
2023-05-05 18:12:26 +00:00
candidates, # type: _Pointer[llama_token_data_array]
2023-05-05 17:54:22 +00:00
tau: c_float,
eta: c_float,
2023-05-05 18:12:26 +00:00
mu, # type: _Pointer[c_float]
) -> int:
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_token_mirostat_v2(ctx, candidates, tau, eta, mu)
_lib.llama_sample_token_mirostat_v2.argtypes = [
llama_context_p,
llama_token_data_array_p,
2023-03-29 01:10:23 +00:00
c_float,
c_float,
2023-05-05 18:04:12 +00:00
c_float_p,
2023-05-01 14:44:28 +00:00
]
_lib.llama_sample_token_mirostat_v2.restype = llama_token
2023-05-01 18:02:06 +00:00
# @details Selects the token with the highest probability.
2023-05-21 21:47:21 +00:00
# LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
def llama_sample_token_greedy(
2023-05-05 18:12:26 +00:00
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
) -> int:
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_token_greedy(ctx, candidates)
_lib.llama_sample_token_greedy.argtypes = [
llama_context_p,
llama_token_data_array_p,
]
_lib.llama_sample_token_greedy.restype = llama_token
2023-05-01 18:02:06 +00:00
# @details Randomly selects a token from the candidates based on their probabilities.
2023-05-21 21:47:21 +00:00
# LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
def llama_sample_token(
2023-05-05 18:12:26 +00:00
ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array]
) -> int:
2023-05-01 14:44:28 +00:00
return _lib.llama_sample_token(ctx, candidates)
_lib.llama_sample_token.argtypes = [
llama_context_p,
llama_token_data_array_p,
]
2023-05-01 14:44:28 +00:00
_lib.llama_sample_token.restype = llama_token
# Performance information
2023-03-24 18:59:29 +00:00
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_print_timings(struct llama_context * ctx);
2023-03-23 09:33:06 +00:00
def llama_print_timings(ctx: llama_context_p):
_lib.llama_print_timings(ctx)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_print_timings.argtypes = [llama_context_p]
_lib.llama_print_timings.restype = None
2023-05-21 21:47:21 +00:00
# LLAMA_API void llama_reset_timings(struct llama_context * ctx);
2023-03-23 09:33:06 +00:00
def llama_reset_timings(ctx: llama_context_p):
_lib.llama_reset_timings(ctx)
2023-03-23 09:33:06 +00:00
2023-03-24 18:35:41 +00:00
_lib.llama_reset_timings.argtypes = [llama_context_p]
_lib.llama_reset_timings.restype = None
# Print system information
2023-05-21 21:47:21 +00:00
# LLAMA_API const char * llama_print_system_info(void);
2023-03-23 09:33:06 +00:00
def llama_print_system_info() -> bytes:
return _lib.llama_print_system_info()
_lib.llama_print_system_info.argtypes = []
_lib.llama_print_system_info.restype = c_char_p
2023-05-21 21:47:21 +00:00
###################################################################################################
_llama_initialized = False
if not _llama_initialized:
llama_init_backend()
_llama_initialized = True