Repository navigation
Expand file tree
/
Copy pathmodel.py
More file actions
367 lines (338 loc) · 15.9 KB
/
Copy pathmodel.py
File metadata and controls
367 lines (338 loc) · 15.9 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
from __future__ import annotations
import json
from typing import Dict, Optional, Union, List
import llama_cpp
import llama_cpp.llama_speculative as llama_speculative
import llama_cpp.llama_tokenizer as llama_tokenizer
from llama_cpp.server.settings import ModelSettings
def _create_prompt_cache(settings: ModelSettings):
if settings.cache_type == "disk":
if settings.verbose:
print(f"Using disk cache with size {settings.cache_size}")
return llama_cpp.LlamaDiskCache(capacity_bytes=settings.cache_size)
if settings.cache_type == "trie":
if settings.verbose:
print(f"Using trie cache with size {settings.cache_size}")
return llama_cpp.LlamaTrieCache(capacity_bytes=settings.cache_size)
if settings.verbose:
print(f"Using ram cache with size {settings.cache_size}")
return llama_cpp.LlamaRAMCache(capacity_bytes=settings.cache_size)
class LlamaProxy:
def __init__(self, models: List[ModelSettings]) -> None:
assert len(models) > 0, "No models provided!"
self._model_settings_dict: dict[str, ModelSettings] = {}
for model in models:
if not model.model_alias:
model.model_alias = model.model
self._model_settings_dict[model.model_alias] = model
self._current_model: Optional[llama_cpp.Llama] = None
self._current_model_alias: Optional[str] = None
self._default_model_settings: ModelSettings = models[0]
self._default_model_alias: str = self._default_model_settings.model_alias # type: ignore
# Load default model
self._current_model = self.load_llama_from_model_settings(
self._default_model_settings
)
self._current_model_alias = self._default_model_alias
def __call__(self, model: Optional[str] = None) -> llama_cpp.Llama:
if model is None:
model = self._default_model_alias
if model not in self._model_settings_dict:
model = self._default_model_alias
if model == self._current_model_alias:
if self._current_model is not None:
return self._current_model
if self._current_model:
self._current_model.close()
self._current_model = None
settings = self._model_settings_dict[model]
self._current_model = self.load_llama_from_model_settings(settings)
self._current_model_alias = model
return self._current_model
def __getitem__(self, model: str):
return self._model_settings_dict[model].model_dump()
def __setitem__(self, model: str, settings: Union[ModelSettings, str, bytes]):
if isinstance(settings, (bytes, str)):
settings = ModelSettings.model_validate_json(settings)
self._model_settings_dict[model] = settings
def __iter__(self):
for model in self._model_settings_dict:
yield model
def free(self):
if self._current_model:
self._current_model.close()
del self._current_model
@staticmethod
def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama:
chat_handler = None
if settings.chat_format == "llava-1-5":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.Llava15ChatHandler.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.Llava15ChatHandler(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "obsidian":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.ObsidianChatHandler.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.ObsidianChatHandler(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "llava-1-6":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.Llava16ChatHandler.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.Llava16ChatHandler(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "moondream":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.MoondreamChatHandler.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.MoondreamChatHandler(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "nanollava":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.NanoLlavaChatHandler.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.NanoLlavaChatHandler(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "llama-3-vision-alpha":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.Llama3VisionAlpha.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.Llama3VisionAlpha(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "minicpm-v-2.6":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.MiniCPMv26ChatHandler.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.MiniCPMv26ChatHandler(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "gemma3":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.Gemma3ChatHandler.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.Gemma3ChatHandler(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "qwen2.5-vl":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.Qwen25VLChatHandler.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.Qwen25VLChatHandler(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "lfm2-vl":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.LFM2VLChatHandler.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.LFM2VLChatHandler(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "lfm2.5-vl":
assert settings.clip_model_path is not None, "clip model not found"
if settings.hf_model_repo_id is not None:
chat_handler = (
llama_cpp.llama_chat_format.LFM25VLChatHandler.from_pretrained(
repo_id=settings.hf_model_repo_id,
filename=settings.clip_model_path,
verbose=settings.verbose,
)
)
else:
chat_handler = llama_cpp.llama_chat_format.LFM25VLChatHandler(
clip_model_path=settings.clip_model_path, verbose=settings.verbose
)
elif settings.chat_format == "hf-autotokenizer":
assert (
settings.hf_pretrained_model_name_or_path is not None
), "hf_pretrained_model_name_or_path must be set for hf-autotokenizer"
chat_handler = (
llama_cpp.llama_chat_format.hf_autotokenizer_to_chat_completion_handler(
settings.hf_pretrained_model_name_or_path
)
)
elif settings.chat_format == "hf-tokenizer-config":
assert (
settings.hf_tokenizer_config_path is not None
), "hf_tokenizer_config_path must be set for hf-tokenizer-config"
chat_handler = llama_cpp.llama_chat_format.hf_tokenizer_config_to_chat_completion_handler(
json.load(open(settings.hf_tokenizer_config_path))
)
tokenizer: Optional[llama_cpp.BaseLlamaTokenizer] = None
if settings.hf_pretrained_model_name_or_path is not None:
tokenizer = llama_tokenizer.LlamaHFTokenizer.from_pretrained(
settings.hf_pretrained_model_name_or_path
)
draft_model = None
if settings.draft_model is not None:
draft_model = llama_speculative.LlamaPromptLookupDecoding(
num_pred_tokens=settings.draft_model_num_pred_tokens
)
kv_overrides: Optional[Dict[str, Union[bool, int, float, str]]] = None
if settings.kv_overrides is not None:
assert isinstance(settings.kv_overrides, list)
kv_overrides = {}
for kv in settings.kv_overrides:
key, value = kv.split("=")
if ":" in value:
value_type, value = value.split(":")
if value_type == "bool":
kv_overrides[key] = value.lower() in ["true", "1"]
elif value_type == "int":
kv_overrides[key] = int(value)
elif value_type == "float":
kv_overrides[key] = float(value)
elif value_type == "str":
kv_overrides[key] = value
else:
raise ValueError(f"Unknown value type {value_type}")
import functools
kwargs = {}
if settings.hf_model_repo_id is not None:
create_fn = functools.partial(
llama_cpp.Llama.from_pretrained,
repo_id=settings.hf_model_repo_id,
filename=settings.model,
)
else:
create_fn = llama_cpp.Llama
kwargs["model_path"] = settings.model
_model = create_fn(
**kwargs,
# Model Params
n_gpu_layers=settings.n_gpu_layers,
split_mode=settings.split_mode,
load_mode=settings.load_mode,
main_gpu=settings.main_gpu,
tensor_split=settings.tensor_split,
vocab_only=settings.vocab_only,
check_tensors=settings.check_tensors,
use_extra_bufts=settings.use_extra_bufts,
no_host=settings.no_host,
kv_overrides=kv_overrides,
rpc_servers=settings.rpc_servers,
rpc_local_devices=settings.rpc_local_devices,
# Context Params
seed=settings.seed,
n_ctx=settings.n_ctx,
n_batch=settings.n_batch,
n_ubatch=settings.n_ubatch,
n_seq_max=settings.n_seq_max,
n_threads=settings.n_threads,
n_threads_batch=settings.n_threads_batch,
rope_scaling_type=settings.rope_scaling_type,
pooling_type=settings.pooling_type,
attention_type=settings.attention_type,
flash_attn_type=settings.flash_attn_type,
rope_freq_base=settings.rope_freq_base,
rope_freq_scale=settings.rope_freq_scale,
yarn_ext_factor=settings.yarn_ext_factor,
yarn_attn_factor=settings.yarn_attn_factor,
yarn_beta_fast=settings.yarn_beta_fast,
yarn_beta_slow=settings.yarn_beta_slow,
yarn_orig_ctx=settings.yarn_orig_ctx,
mul_mat_q=settings.mul_mat_q,
logits_all=settings.logits_all,
embeddings=settings.embeddings,
offload_kqv=settings.offload_kqv,
no_perf=settings.no_perf,
op_offload=settings.op_offload,
swa_full=settings.swa_full,
kv_unified=settings.kv_unified,
# Sampling Params
last_n_tokens_size=settings.last_n_tokens_size,
# Backend Params
numa=settings.numa,
# Chat Format Params
chat_format=settings.chat_format,
chat_handler=chat_handler,
# Speculative Decoding
draft_model=draft_model,
# KV Cache Quantization
type_k=settings.type_k,
type_v=settings.type_v,
# Tokenizer
tokenizer=tokenizer,
# Misc
verbose=settings.verbose,
)
if settings.cache:
_model.set_cache(_create_prompt_cache(settings))
return _model