llama.cpp/llama_cpp/llama_cpp.py

224 lines
5.9 KiB
Python
Raw Normal View History

2023-03-23 09:33:06 +00:00
import ctypes
2023-03-24 18:35:41 +00:00
from ctypes import (
c_int,
c_float,
c_double,
c_char_p,
c_void_p,
c_bool,
POINTER,
Structure,
)
2023-03-23 09:33:06 +00:00
import pathlib
# Load the library
2023-03-23 17:54:14 +00:00
libfile = pathlib.Path(__file__).parent / "libllama.so"
2023-03-23 09:33:06 +00:00
lib = ctypes.CDLL(str(libfile))
# C types
llama_token = c_int
llama_token_p = POINTER(llama_token)
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
class llama_token_data(Structure):
_fields_ = [
2023-03-24 18:35:41 +00:00
("id", llama_token), # token id
("p", c_float), # probability of the token
("plog", c_float), # log probability of the token
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)
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
class llama_context_params(Structure):
_fields_ = [
2023-03-24 18:35:41 +00:00
("n_ctx", c_int), # text context
("n_parts", c_int), # -1 for default
("seed", c_int), # RNG seed, 0 for random
("f16_kv", c_bool), # use fp16 for KV cache
(
"logits_all",
c_bool,
), # the llama_eval() call computes all logits, not just the last one
("vocab_only", c_bool), # only load the vocabulary, no weights
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)
llama_context_p = c_void_p
# C functions
lib.llama_context_default_params.argtypes = []
lib.llama_context_default_params.restype = llama_context_params
lib.llama_init_from_file.argtypes = [c_char_p, llama_context_params]
lib.llama_init_from_file.restype = llama_context_p
lib.llama_free.argtypes = [llama_context_p]
lib.llama_free.restype = None
lib.llama_model_quantize.argtypes = [c_char_p, c_char_p, c_int, c_int]
lib.llama_model_quantize.restype = c_int
lib.llama_eval.argtypes = [llama_context_p, llama_token_p, c_int, c_int, c_int]
lib.llama_eval.restype = c_int
lib.llama_tokenize.argtypes = [llama_context_p, c_char_p, llama_token_p, c_int, c_bool]
lib.llama_tokenize.restype = c_int
lib.llama_n_vocab.argtypes = [llama_context_p]
lib.llama_n_vocab.restype = c_int
lib.llama_n_ctx.argtypes = [llama_context_p]
lib.llama_n_ctx.restype = c_int
lib.llama_get_logits.argtypes = [llama_context_p]
lib.llama_get_logits.restype = POINTER(c_float)
lib.llama_token_to_str.argtypes = [llama_context_p, llama_token]
lib.llama_token_to_str.restype = c_char_p
lib.llama_token_bos.argtypes = []
lib.llama_token_bos.restype = llama_token
lib.llama_token_eos.argtypes = []
lib.llama_token_eos.restype = llama_token
2023-03-24 18:35:41 +00:00
lib.llama_sample_top_p_top_k.argtypes = [
llama_context_p,
llama_token_p,
c_int,
c_int,
c_double,
c_double,
c_double,
]
2023-03-23 09:33:06 +00:00
lib.llama_sample_top_p_top_k.restype = llama_token
lib.llama_print_timings.argtypes = [llama_context_p]
lib.llama_print_timings.restype = None
lib.llama_reset_timings.argtypes = [llama_context_p]
lib.llama_reset_timings.restype = None
lib.llama_print_system_info.argtypes = []
lib.llama_print_system_info.restype = c_char_p
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
# Python functions
def llama_context_default_params() -> llama_context_params:
params = lib.llama_context_default_params()
return params
2023-03-24 18:35:41 +00:00
def llama_init_from_file(
path_model: bytes, params: llama_context_params
) -> llama_context_p:
2023-03-23 09:33:06 +00:00
"""Various functions for loading a ggml llama model.
Allocate (almost) all memory needed for the model.
2023-03-24 18:35:41 +00:00
Return NULL on failure"""
2023-03-23 09:33:06 +00:00
return lib.llama_init_from_file(path_model, params)
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
def llama_free(ctx: llama_context_p):
"""Free all allocated memory"""
lib.llama_free(ctx)
2023-03-24 18:35:41 +00:00
def llama_model_quantize(
fname_inp: bytes, fname_out: bytes, itype: c_int, qk: c_int
) -> c_int:
2023-03-23 09:33:06 +00:00
"""Returns 0 on success"""
return lib.llama_model_quantize(fname_inp, fname_out, itype, qk)
2023-03-24 18:35:41 +00:00
def llama_eval(
ctx: llama_context_p,
tokens: llama_token_p,
n_tokens: c_int,
n_past: c_int,
n_threads: c_int,
) -> c_int:
2023-03-23 09:33:06 +00:00
"""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"""
return lib.llama_eval(ctx, tokens, n_tokens, n_past, n_threads)
2023-03-24 18:35:41 +00:00
def llama_tokenize(
ctx: llama_context_p,
text: bytes,
tokens: llama_token_p,
n_max_tokens: c_int,
add_bos: c_bool,
) -> c_int:
2023-03-23 09:33:06 +00:00
"""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
2023-03-24 18:35:41 +00:00
Returns a negative number on failure - the number of tokens that would have been returned
"""
2023-03-23 09:33:06 +00:00
return lib.llama_tokenize(ctx, text, tokens, n_max_tokens, add_bos)
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
def llama_n_vocab(ctx: llama_context_p) -> c_int:
return lib.llama_n_vocab(ctx)
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
def llama_n_ctx(ctx: llama_context_p) -> c_int:
return lib.llama_n_ctx(ctx)
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
def llama_get_logits(ctx: llama_context_p):
"""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"""
return lib.llama_get_logits(ctx)
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
def llama_token_to_str(ctx: llama_context_p, token: int) -> bytes:
"""Token Id -> String. Uses the vocabulary in the provided context"""
return lib.llama_token_to_str(ctx, token)
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
def llama_token_bos() -> llama_token:
return lib.llama_token_bos()
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
def llama_token_eos() -> llama_token:
return lib.llama_token_eos()
2023-03-24 18:35:41 +00:00
def llama_sample_top_p_top_k(
ctx: llama_context_p,
last_n_tokens_data: llama_token_p,
last_n_tokens_size: c_int,
top_k: c_int,
top_p: c_double,
temp: c_double,
repeat_penalty: c_double,
) -> llama_token:
return lib.llama_sample_top_p_top_k(
ctx, last_n_tokens_data, last_n_tokens_size, top_k, top_p, temp, repeat_penalty
)
2023-03-23 09:33:06 +00:00
def llama_print_timings(ctx: llama_context_p):
lib.llama_print_timings(ctx)
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
def llama_reset_timings(ctx: llama_context_p):
lib.llama_reset_timings(ctx)
2023-03-24 18:35:41 +00:00
2023-03-23 09:33:06 +00:00
def llama_print_system_info() -> bytes:
"""Print system informaiton"""
return lib.llama_print_system_info()