forked from abetlen/llama-cpp-python
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathllama_cpp.py
More file actions
380 lines (257 loc) · 10.4 KB
/
llama_cpp.py
File metadata and controls
380 lines (257 loc) · 10.4 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
import sys
import os
import ctypes
from ctypes import (
c_int,
c_float,
c_char_p,
c_void_p,
c_bool,
POINTER,
Structure,
Array,
c_uint8,
c_size_t,
)
import pathlib
# Load the library
def _load_shared_library(lib_base_name):
# Determine the file extension based on the platform
if sys.platform.startswith("linux"):
lib_ext = ".so"
elif sys.platform == "darwin":
lib_ext = ".so"
elif sys.platform == "win32":
lib_ext = ".dll"
else:
raise RuntimeError("Unsupported platform")
# 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 = [
_base_path / f"lib{lib_base_name}{lib_ext}",
_base_path / f"{lib_base_name}{lib_ext}",
]
if "LLAMA_CPP_LIB" in os.environ:
lib_base_name = os.environ["LLAMA_CPP_LIB"]
_lib = pathlib.Path(lib_base_name)
_base_path = _lib.parent.resolve()
_lib_paths = [_lib.resolve()]
# 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))
# 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))
except Exception as e:
raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}")
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)
# C types
llama_context_p = c_void_p
llama_token = c_int
llama_token_p = POINTER(llama_token)
class llama_token_data(Structure):
_fields_ = [
("id", llama_token), # token id
("p", c_float), # probability of the token
("plog", c_float), # log probability of the token
]
llama_token_data_p = POINTER(llama_token_data)
llama_progress_callback = ctypes.CFUNCTYPE(None, c_float, c_void_p)
class llama_context_params(Structure):
_fields_ = [
("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
("use_mmap", c_bool), # use mmap if possible
("use_mlock", c_bool), # force system to keep model in RAM
("embedding", c_bool), # embedding mode only
# called with a progress value between 0 and 1, pass NULL to disable
("progress_callback", llama_progress_callback),
# context pointer passed to the progress callback
("progress_callback_user_data", c_void_p),
]
llama_context_params_p = POINTER(llama_context_params)
LLAMA_FTYPE_ALL_F32 = ctypes.c_int(0)
LLAMA_FTYPE_MOSTLY_F16 = ctypes.c_int(1) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0 = ctypes.c_int(2) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1 = ctypes.c_int(3) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = ctypes.c_int(4) # tok_embeddings.weight and output.weight are F16
# Functions
def llama_context_default_params() -> llama_context_params:
return _lib.llama_context_default_params()
_lib.llama_context_default_params.argtypes = []
_lib.llama_context_default_params.restype = llama_context_params
def llama_mmap_supported() -> c_bool:
return _lib.llama_mmap_supported()
_lib.llama_mmap_supported.argtypes = []
_lib.llama_mmap_supported.restype = c_bool
def llama_mlock_supported() -> c_bool:
return _lib.llama_mlock_supported()
_lib.llama_mlock_supported.argtypes = []
_lib.llama_mlock_supported.restype = c_bool
# Various functions for loading a ggml llama model.
# Allocate (almost) all memory needed for the model.
# Return NULL on failure
def llama_init_from_file(
path_model: bytes, params: llama_context_params
) -> llama_context_p:
return _lib.llama_init_from_file(path_model, params)
_lib.llama_init_from_file.argtypes = [c_char_p, llama_context_params]
_lib.llama_init_from_file.restype = llama_context_p
# Frees all allocated memory
def llama_free(ctx: llama_context_p):
_lib.llama_free(ctx)
_lib.llama_free.argtypes = [llama_context_p]
_lib.llama_free.restype = None
# TODO: not great API - very likely to change
# Returns 0 on success
def llama_model_quantize(fname_inp: bytes, fname_out: bytes, itype: c_int) -> c_int:
return _lib.llama_model_quantize(fname_inp, fname_out, itype)
_lib.llama_model_quantize.argtypes = [c_char_p, c_char_p, c_int]
_lib.llama_model_quantize.restype = c_int
# Returns the KV cache that will contain the context for the
# ongoing prediction with the model.
def llama_get_kv_cache(ctx: llama_context_p):
return _lib.llama_get_kv_cache(ctx)
_lib.llama_get_kv_cache.argtypes = [llama_context_p]
_lib.llama_get_kv_cache.restype = POINTER(c_uint8)
# Returns the size of the KV cache
def llama_get_kv_cache_size(ctx: llama_context_p) -> c_size_t:
return _lib.llama_get_kv_cache_size(ctx)
_lib.llama_get_kv_cache_size.argtypes = [llama_context_p]
_lib.llama_get_kv_cache_size.restype = c_size_t
# Returns the number of tokens in the KV cache
def llama_get_kv_cache_token_count(ctx: llama_context_p) -> c_int:
return _lib.llama_get_kv_cache_token_count(ctx)
_lib.llama_get_kv_cache_token_count.argtypes = [llama_context_p]
_lib.llama_get_kv_cache_token_count.restype = c_int
# Sets the KV cache containing the current context for the model
def llama_set_kv_cache(
ctx: llama_context_p, kv_cache, n_size: c_size_t, n_token_count: c_int
):
return _lib.llama_set_kv_cache(ctx, kv_cache, n_size, n_token_count)
_lib.llama_set_kv_cache.argtypes = [llama_context_p, POINTER(c_uint8), c_size_t, c_int]
_lib.llama_set_kv_cache.restype = None
# 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
def llama_eval(
ctx: llama_context_p,
tokens, # type: Array[llama_token]
n_tokens: c_int,
n_past: c_int,
n_threads: c_int,
) -> c_int:
return _lib.llama_eval(ctx, tokens, n_tokens, n_past, n_threads)
_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
def llama_tokenize(
ctx: llama_context_p,
text: bytes,
tokens, # type: Array[llama_token]
n_max_tokens: c_int,
add_bos: c_bool,
) -> c_int:
return _lib.llama_tokenize(ctx, text, tokens, n_max_tokens, add_bos)
_lib.llama_tokenize.argtypes = [llama_context_p, c_char_p, llama_token_p, c_int, c_bool]
_lib.llama_tokenize.restype = c_int
def llama_n_vocab(ctx: llama_context_p) -> c_int:
return _lib.llama_n_vocab(ctx)
_lib.llama_n_vocab.argtypes = [llama_context_p]
_lib.llama_n_vocab.restype = c_int
def llama_n_ctx(ctx: llama_context_p) -> c_int:
return _lib.llama_n_ctx(ctx)
_lib.llama_n_ctx.argtypes = [llama_context_p]
_lib.llama_n_ctx.restype = c_int
def llama_n_embd(ctx: llama_context_p) -> c_int:
return _lib.llama_n_embd(ctx)
_lib.llama_n_embd.argtypes = [llama_context_p]
_lib.llama_n_embd.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
def llama_get_logits(ctx: llama_context_p):
return _lib.llama_get_logits(ctx)
_lib.llama_get_logits.argtypes = [llama_context_p]
_lib.llama_get_logits.restype = POINTER(c_float)
# Get the embeddings for the input
# shape: [n_embd] (1-dimensional)
def llama_get_embeddings(ctx: llama_context_p):
return _lib.llama_get_embeddings(ctx)
_lib.llama_get_embeddings.argtypes = [llama_context_p]
_lib.llama_get_embeddings.restype = POINTER(c_float)
# Token Id -> String. Uses the vocabulary in the provided context
def llama_token_to_str(ctx: llama_context_p, token: llama_token) -> bytes:
return _lib.llama_token_to_str(ctx, token)
_lib.llama_token_to_str.argtypes = [llama_context_p, llama_token]
_lib.llama_token_to_str.restype = c_char_p
# Special tokens
def llama_token_bos() -> llama_token:
return _lib.llama_token_bos()
_lib.llama_token_bos.argtypes = []
_lib.llama_token_bos.restype = llama_token
def llama_token_eos() -> llama_token:
return _lib.llama_token_eos()
_lib.llama_token_eos.argtypes = []
_lib.llama_token_eos.restype = llama_token
# TODO: improve the last_n_tokens interface ?
def llama_sample_top_p_top_k(
ctx: llama_context_p,
last_n_tokens_data, # type: Array[llama_token]
last_n_tokens_size: c_int,
top_k: c_int,
top_p: c_float,
temp: c_float,
repeat_penalty: c_float,
) -> 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
)
_lib.llama_sample_top_p_top_k.argtypes = [
llama_context_p,
llama_token_p,
c_int,
c_int,
c_float,
c_float,
c_float,
]
_lib.llama_sample_top_p_top_k.restype = llama_token
# Performance information
def llama_print_timings(ctx: llama_context_p):
_lib.llama_print_timings(ctx)
_lib.llama_print_timings.argtypes = [llama_context_p]
_lib.llama_print_timings.restype = None
def llama_reset_timings(ctx: llama_context_p):
_lib.llama_reset_timings(ctx)
_lib.llama_reset_timings.argtypes = [llama_context_p]
_lib.llama_reset_timings.restype = None
# Print system information
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