forked from ggerganov/llama.cpp
-
Notifications
You must be signed in to change notification settings - Fork 364
/
class.py
336 lines (307 loc) · 16.4 KB
/
class.py
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
## KoboldCpp based GGML Backend by Concedo
## For use as a custom backend in KoboldAI United
## Not intended for general use.
from __future__ import annotations
import time, json
import torch
import requests
import numpy as np
from typing import List, Optional, Union
import os, time
from . import koboldcpp
import utils
from logger import logger
from modeling.inference_model import (
GenerationResult,
GenerationSettings,
InferenceModel,
)
model_backend_name = "KoboldCPP" #specific instead of ggml
model_backend_type = "ggml" #This should be a generic name in case multiple model backends are compatible (think Hugging Face Custom and Basic Hugging Face)
class KoboldCppException(Exception):
"""To be used for errors on cpp side of KoboldCpp."""
class KcppArgsObject:
def __init__(self, **kwargs):
self.__dict__.update(kwargs)
class model_backend(InferenceModel):
def __init__(self) -> None:
super().__init__()
self.kcpp_backend_loaded = False
def is_valid(self, model_name, model_path, menu_path):
foundfile = False
try:
files = os.listdir(model_path)
foundfile = len([filename for filename in files if (("ggml" in filename.lower() and ".bin" in filename.lower()) or ".gguf" in filename.lower())])>0
except:
pass
return foundfile
def get_requested_parameters(self, model_name, model_path, menu_path, parameters = {}):
self.kcpp_threads = 5
self.model_name = "GGML_Model"
self.kcpp_ctxsize = 2048
self.kcpp_blasbatchsize = 512
self.kcpp_gpulayers = 0
self.kcpp_smartcontext = False
self.kcpp_ropescale = 0.0
self.kcpp_ropebase = 10000.0
self.kcpp_useclblast = None
self.kcpp_usecublas = None
self.kcpp_usecpu = False
self.kcpp_noavx2 = False
self.kcpp_nommap = False
self.kcpp_usevulkan = None
self.kcpp_debugmode = 0
self.kcpp_tensor_split_str = ""
self.kcpp_tensor_split = None
files = os.listdir(model_path)
foundfiles = [filename for filename in files if (("ggml" in filename.lower() and ".bin" in filename.lower()) or ".gguf" in filename.lower())]
requested_parameters = []
foldermdls = []
for ff in foundfiles:
foldermdls.append({'text': ff, 'value': os.path.join(model_path, ff)})
requested_parameters.append({
"uitype": "dropdown",
"unit": "string",
"label": "GGML DataFile Name",
"id": "kcpp_filename",
"default": os.path.join(model_path, foundfiles[0]) if len(foundfiles)>0 else model_name,
"check": {"value": "", 'check': "!="},
"tooltip": "Actual GGML DataFile Name",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": "",
'children': foldermdls
})
requested_parameters.append({
"uitype": "dropdown",
"unit": "int",
"label": "KoboldCpp Accelerator",
"id": "kcpp_accelerator",
"default": 0,
"check": {"value": "", 'check': "!="},
'multiple': False,
"tooltip": "KoboldCpp Accelerator",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": "",
'children': [{'text': 'Use No BLAS', 'value': 0}, {'text': 'Use CuBLAS', 'value': 1},
{'text': 'Use CLBLast GPU #1', 'value': 2},{'text': 'Use CLBLast GPU #2', 'value': 3},{'text': 'Use CLBLast GPU #3', 'value': 4}
,{'text': 'NoAVX2 Mode (Old CPU)', 'value': 5},{'text': 'Failsafe Mode (Old CPU)', 'value': 6},{'text': 'Use Vulkan GPU #1', 'value': 7},{'text': 'Use Vulkan GPU #2', 'value': 8}],
})
requested_parameters.append({
"uitype": "text",
"unit": "int",
"label": "Threads",
"id": "kcpp_threads",
"default": self.kcpp_threads,
"check": {"value": "", 'check': "!="},
"tooltip": "Thread Count",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": ""
})
requested_parameters.append({
"uitype": "text",
"unit": "int",
"label": "Max Context Size",
"id": "kcpp_ctxsize",
"default": self.kcpp_ctxsize,
"check": {"value": "", 'check': "!="},
"tooltip": "Max Context Size",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": ""
})
requested_parameters.append({
"uitype": "text",
"unit": "int",
"label": "BLAS Batch Size",
"id": "kcpp_blasbatchsize",
"default": self.kcpp_blasbatchsize,
"check": {"value": "", 'check': "!="},
"tooltip": "BLAS Batch Size",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": ""
})
requested_parameters.append({
"uitype": "text",
"unit": "int",
"label": "GPU Layers",
"id": "kcpp_gpulayers",
"default": self.kcpp_gpulayers,
"check": {"value": "", 'check': "!="},
"tooltip": "GPU Layers",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": ""
})
requested_parameters.append({
"uitype": "text",
"unit": "int",
"label": "Rope Scale",
"id": "kcpp_ropescale",
"default": self.kcpp_ropescale,
"check": {"value": "", 'check': "!="},
"tooltip": "Rope Scale",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": ""
})
requested_parameters.append({
"uitype": "text",
"unit": "int",
"label": "Rope Base",
"id": "kcpp_ropebase",
"default": self.kcpp_ropebase,
"check": {"value": "", 'check': "!="},
"tooltip": "Rope Base",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": ""
})
requested_parameters.append({
"uitype": "dropdown",
"unit": "int",
"label": "Smart Context",
"id": "kcpp_smartcontext",
"default": self.kcpp_smartcontext,
"check": {"value": "", 'check': "!="},
'multiple': False,
"tooltip": "Smart Context",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": "",
'children': [{'text': 'False', 'value': False}, {'text': 'True', 'value': True}],
})
requested_parameters.append({
"uitype": "text",
"unit": "text",
"label": "GPU ID",
"id": "kcpp_tensor_split_str",
"default": "1",
"check": {"value": "", 'check': "!="},
"tooltip": "Which GPU's do we use? For example:1 2",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": ""
})
requested_parameters.append({
"uitype": "dropdown",
"unit": "int",
"label": "Debug Mode",
"id": "kcpp_debugmode",
"default": self.kcpp_debugmode,
"check": {"value": "", 'check': "!="},
'multiple': False,
"tooltip": "Debug Mode",
"menu_path": "",
"refresh_model_inputs": False,
"extra_classes": "",
'children': [{'text': 'False', 'value': 0}, {'text': 'True', 'value': 1}],
})
return requested_parameters
def set_input_parameters(self, parameters):
self.kcpp_threads = parameters["kcpp_threads"]
self.kcpp_filename = parameters["kcpp_filename"]
self.kcpp_ctxsize = parameters["kcpp_ctxsize"]
self.kcpp_blasbatchsize = parameters["kcpp_blasbatchsize"]
self.kcpp_gpulayers = parameters["kcpp_gpulayers"]
self.kcpp_smartcontext = parameters["kcpp_smartcontext"]
self.kcpp_ropescale = parameters["kcpp_ropescale"]
self.kcpp_ropebase = parameters["kcpp_ropebase"]
self.kcpp_debugmode = parameters["kcpp_debugmode"]
self.kcpp_tensor_split_str = parameters["kcpp_tensor_split_str"]
if self.kcpp_tensor_split_str and self.kcpp_tensor_split_str!="":
splits = self.kcpp_tensor_split_str.split()
self.kcpp_tensor_split = []
for s in splits:
self.kcpp_tensor_split.append(int(s))
print(self.kcpp_tensor_split)
accel = parameters["kcpp_accelerator"]
if accel==0:
self.kcpp_usecpu = True
elif accel==1:
self.kcpp_usecublas = ["normal"]
elif accel==2:
self.kcpp_useclblast = [0,0]
elif accel==3:
self.kcpp_useclblast = [1,0]
elif accel==4:
self.kcpp_useclblast = [0,1]
elif accel==5:
self.kcpp_noavx2 = True
elif accel==6:
self.kcpp_noavx2 = True
self.kcpp_usecpu = True
self.kcpp_nommap = True
elif accel==7:
self.kcpp_usevulkan = [0]
elif accel==8:
self.kcpp_usevulkan = [1]
pass
def unload(self):
print("Attemping to unload library")
self.process.terminate()
def _load(self, save_model: bool, initial_load: bool) -> None:
self.tokenizer = self._get_tokenizer("gpt2")
kcppargs = KcppArgsObject(model=self.kcpp_filename, model_param=self.kcpp_filename,
port=5001, port_param=5001, host='', launch=False, lora=None, threads=self.kcpp_threads, blasthreads=self.kcpp_threads,
psutil_set_threads=False, highpriority=False, contextsize=self.kcpp_ctxsize, blasbatchsize=self.kcpp_blasbatchsize,
ropeconfig=[self.kcpp_ropescale, self.kcpp_ropebase], stream=False, smartcontext=self.kcpp_smartcontext, forceversion=0,
nommap=self.kcpp_nommap, usemlock=False, noavx2=self.kcpp_noavx2, debugmode=self.kcpp_debugmode, skiplauncher=True, usecpu=self.kcpp_usecpu,
useclblast=self.kcpp_useclblast, usecublas=self.kcpp_usecublas, usevulkan=self.kcpp_usevulkan, gpulayers=self.kcpp_gpulayers,
tensor_split=self.kcpp_tensor_split, config=None, onready='', multiuser=False, foreground=False, preloadstory=None, noshift=False,
remotetunnel=False, ssl=False, benchmark=None, nocertify=False, mmproj=None, password=None, chatcompletionsadapter=None)
#koboldcpp.main(kcppargs,False) #initialize library without enabling Lite http server
(self.output_queue, self.input_queue, self.process) = koboldcpp.start_in_seperate_process(kcppargs)
while True:
data = self.output_queue.get()
if data['command'] == 'load status':
utils.koboldai_vars.total_layers = data['data']['total']
utils.koboldai_vars.loaded_layers = data['data']['loaded']
elif data['command'] == 'complete':
break
time.sleep(0.02)
def _save_settings(self):
pass
def _raw_generate(
self,
prompt_tokens: Union[List[int], torch.Tensor],
max_new: int,
gen_settings: GenerationSettings,
single_line: bool = False,
batch_count: int = 1,
seed: Optional[int] = None,
**kwargs,
) -> GenerationResult:
decoded_prompt = utils.decodenewlines(self.tokenizer.decode(prompt_tokens))
# Store context in memory to use it for comparison with generated content
utils.koboldai_vars.lastctx = decoded_prompt
self.input_queue.put({'command': 'generate', 'data': {'prompt':decoded_prompt, 'max_length': max_new, 'max_context_length': utils.koboldai_vars.max_length,
'temperature': gen_settings.temp, 'top_k': int(gen_settings.top_k), 'top_a': gen_settings.top_a, 'top_p': gen_settings.top_p,
'typical_p': gen_settings.typical, 'tfs': gen_settings.tfs, 'rep_pen': gen_settings.rep_pen, 'rep_pen_range': gen_settings.rep_pen_range,
"sampler_order": gen_settings.sampler_order, "use_default_badwordsids": utils.koboldai_vars.use_default_badwordsids}
})
#genresult = koboldcpp.generate(decoded_prompt,"",max_new,utils.koboldai_vars.max_length,
#gen_settings.temp,int(gen_settings.top_k),gen_settings.top_a,gen_settings.top_p,
#gen_settings.typical,gen_settings.tfs,gen_settings.rep_pen,gen_settings.rep_pen_range,
#sampler_order=gen_settings.sampler_order,use_default_badwordsids=utils.koboldai_vars.use_default_badwordsids)
genresult = []
while True:
data = self.output_queue.get()
print(data)
if data['command'] == 'generated text':
genresult.append(data['data'])
if self.output_queue.empty():
break
time.sleep(0.02)
return GenerationResult(
model=self,
out_batches=np.array(
[self.tokenizer.encode(x) for x in genresult]
),
prompt=prompt_tokens,
is_whole_generation=True,
single_line=single_line,
)