-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcd_model.py
317 lines (286 loc) · 14.9 KB
/
cd_model.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
#!/usr/bin/env python
# coding=utf-8
from peft import PeftModel
##
import torch
import copy
class CDModel:
def __init__(self, pretrain_model, LLM,input_ids):
### for CD
self.pretrain_model = pretrain_model
import gol
model_args = gol.get_value('gol_model_args')
tokenizer = gol.get_value('gol_tokenizer')
test_file = gol.get_value('gol_test_file')
self.model_args = model_args
self.tokenizer = tokenizer
if model_args.contrastive_decoding in ['ACD','ACD_CD']:
self.model_cd = PeftModel.from_pretrained(pretrain_model, model_args.peft_path,adapter_name='coupling')
self.model_cd.load_adapter(model_args.peft_path.replace('coupling','identify').replace('1000',str(model_args.sub_step)), adapter_name='identify')
self.model_cd.load_adapter(model_args.peft_path.replace('coupling','classify').replace('1000',str(model_args.sub_step)), adapter_name='classify')
self.alter_pattern = ['ide' for i in range(input_ids.shape[0])]
## for plot
# data_plot = {'cou':[],'ide':[],'cla':[],'next_tokens':[]}
import json
with open(test_file, 'r', encoding="utf-8") as f:
for line in f.readlines():
a = json.loads(line)
self.answer_choices = a['answer_choices']
break
self.cla_tokens = tokenizer.encode("".join(self.answer_choices))[1:-2] # qwen 不一样 TODO
elif model_args.contrastive_decoding=='CD':
self.model_cd = PeftModel.from_pretrained(pretrain_model, model_args.peft_path,adapter_name='coupling')
if '1000' in model_args.peft_path:
self.model_cd.load_adapter(model_args.peft_path.replace('1000','500'), adapter_name='coupling_small')
elif '3000' in model_args.peft_path:
self.model_cd.load_adapter(model_args.peft_path.replace('3000','2000'), adapter_name='coupling_small')
elif model_args.contrastive_decoding in ['CAD']:
import json
with open(test_file, 'r',encoding='utf-8') as f:
for line in f.readlines():
a = json.loads(line)
self.answer_choices = a['answer_choices']
break
def jensen_shannon_divergence(self,p, q):
import torch.nn.functional as F
p = p.float()
q = q.float()
m = 0.5 * (F.softmax(p,dim=-1) + F.softmax(q,dim=-1))
kl_p = F.kl_div(F.log_softmax(p,dim=-1), m, reduction='none').sum(-1)
kl_q = F.kl_div(F.log_softmax(q,dim=-1), m, reduction='none').sum(-1)
jsd = 0.5 * (kl_p + kl_q)
return jsd
def decoding(self, input_ids, model_kwargs, output_attentions, output_hidden_states):
if self.model_args.contrastive_decoding=='normal':
# prepare model inputs
model_inputs = self.pretrain_model.prepare_inputs_for_generation(input_ids, **model_kwargs)
# forward pass to get next token
outputs = self.pretrain_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
next_token_logits = outputs.logits[:, -1, :]
elif self.model_args.contrastive_decoding=='DoLa':
# prepare model inputs
model_inputs = self.pretrain_model.prepare_inputs_for_generation(input_ids, **model_kwargs)
# forward pass to get next token
outputs = self.pretrain_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=True,
)
dict_outputs = {}
for i, early_exit_layer in enumerate([0,2,4,6,8,10,12,14,28]):
lm_logits = self.pretrain_model.lm_head(outputs.hidden_states[early_exit_layer]).permute(1, 0, 2).contiguous()
dict_outputs[early_exit_layer] = lm_logits
jsd_max = [0] * input_ids.shape[0]
layer_max = [0] * input_ids.shape[0]
logits_mature = outputs.logits[:, -1, :]
# print(dict_outputs[0].size()) torch.Size([8, 93, 130528])
for layer, outputs_early_exit in dict_outputs.items():
jsd_now = self.jensen_shannon_divergence(outputs_early_exit[:, -1, :], logits_mature)
for i in range(len(jsd_max)):
if jsd_max[i] < jsd_now[i]:
jsd_max[i] = jsd_now[i]
layer_max[i] = layer
# print(layer_max) #
logits_early_exit = []
for i in range(len(layer_max)): #bach size
logits_early_exit.append(dict_outputs[layer_max[i]][i, -1, :])
logits_early_exit = torch.stack(logits_early_exit,dim=0)
next_token_logits = (1 + self.model_args.alpha) * logits_mature - self.model_args.alpha * logits_early_exit
next_token_logits[logits_mature < self.model_args.max_rate * torch.max(logits_mature)] = -1000
elif self.model_args.contrastive_decoding=='CD':
# prepare model inputs
model_inputs = self.pretrain_model.prepare_inputs_for_generation(input_ids, **model_kwargs)
# forward pass to get next token
self.model_cd.set_adapter('coupling')
outputs_cou = self.model_cd.base_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
self.model_cd.set_adapter('coupling_small')
outputs_cou_500 = self.model_cd.base_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
logits_cou = outputs_cou.logits[:, -1, :] # torch.Size([8, 130528])
logits_cou_500 = outputs_cou_500.logits[:, -1, :] # torch.Size([8, 130528])
next_token_logits = (1 + self.model_args.alpha) * logits_cou - self.model_args.alpha * logits_cou_500
next_token_logits[logits_cou < self.model_args.max_rate * torch.max(logits_cou)] = -1000
outputs = outputs_cou
elif self.model_args.contrastive_decoding in ['CAD']:
# prepare model inputs
model_inputs = self.pretrain_model.prepare_inputs_for_generation(input_ids, **model_kwargs)
input_ids_cd = copy.deepcopy(input_ids)
input_prompts = self.tokenizer.batch_decode(input_ids_cd)
select_input_prompts = []
for pr in input_prompts:
#### CAD
if self.model_args.contrastive_decoding=='CAD':
for ans in self.answer_choices:
pr = pr.replace(ans,"")
select_input_prompts.append(pr)
########
select_input_ids = self.tokenizer(select_input_prompts)
for i in range(input_ids_cd.shape[0]):
sii = select_input_ids['input_ids'][i]
for j in range(len(input_ids_cd[i]) - len(sii)):
input_ids_cd[i][j] = 3
input_ids_cd[i][-len(sii):] = torch.LongTensor(sii)
model_inputs_cd = self.pretrain_model.prepare_inputs_for_generation(input_ids_cd, **model_kwargs)
# forward pass to get next token
outputs = self.pretrain_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
outputs_cd = self.pretrain_model(
**model_inputs_cd,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
logits_cou = outputs.logits[:, -1, :] # torch.Size([8, 130528])
logits_cd = outputs_cd.logits[:, -1, :] # torch.Size([8, 130528])
next_token_logits = (1 + self.model_args.alpha) * logits_cou - self.model_args.alpha * logits_cd
next_token_logits[logits_cou < self.model_args.max_rate * torch.max(logits_cou)] = -1000
elif self.model_args.contrastive_decoding=='ACD':
input_ids_cd = copy.deepcopy(input_ids)
input_prompts = self.tokenizer.batch_decode(input_ids_cd)
select_input_prompts = []
for pr in input_prompts:
select_input_prompts.append(pr[-1]) # CFG
select_input_ids = self.tokenizer(select_input_prompts)
for i in range(input_ids_cd.shape[0]):
sii = select_input_ids['input_ids'][i]
for j in range(len(input_ids_cd[i]) - len(sii)):
input_ids_cd[i][j] = 3
input_ids_cd[i][-len(sii):] = torch.LongTensor(sii)
model_inputs_cd = self.pretrain_model.prepare_inputs_for_generation(input_ids_cd, **model_kwargs)
self.model_cd.set_adapter('coupling')
outputs_cd = self.pretrain_model(
**model_inputs_cd,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
logits_cd = outputs_cd.logits[:, -1, :]
# prepare model inputs
model_inputs = self.pretrain_model.prepare_inputs_for_generation(input_ids, **model_kwargs)
# forward pass to get next token
if 'coupling' not in self.model_args.peft_path:
raise
self.model_cd.set_adapter('coupling')
outputs_cou = self.model_cd.base_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
self.model_cd.set_adapter('classify')
outputs_cla = self.model_cd.base_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
self.model_cd.set_adapter('identify')
outputs_ide = self.model_cd.base_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
logits_cou = outputs_cou.logits[:, -1, :] # torch.Size([8, 130528])
logits_ide = outputs_ide.logits[:, -1, :]
logits_cla = outputs_cla.logits[:, -1, :]
next_token_logits = logits_cla
outputs = outputs_cla
elif self.model_args.contrastive_decoding=='ACD_CD':
input_ids_cd = copy.deepcopy(input_ids)
input_prompts = self.tokenizer.batch_decode(input_ids_cd)
select_input_prompts = []
for pr in input_prompts:
select_input_prompts.append(pr[-1]) # CFG
select_input_ids = self.tokenizer(select_input_prompts)
for i in range(input_ids_cd.shape[0]):
sii = select_input_ids['input_ids'][i]
for j in range(len(input_ids_cd[i]) - len(sii)):
input_ids_cd[i][j] = 3
input_ids_cd[i][-len(sii):] = torch.LongTensor(sii)
model_inputs_cd = self.pretrain_model.prepare_inputs_for_generation(input_ids_cd, **model_kwargs)
self.model_cd.set_adapter('coupling')
outputs_cd = self.pretrain_model(
**model_inputs_cd,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
logits_cd = outputs_cd.logits[:, -1, :]
# prepare model inputs
model_inputs = self.pretrain_model.prepare_inputs_for_generation(input_ids, **model_kwargs)
# forward pass to get next token
if 'coupling' not in self.model_args.peft_path:
raise
self.model_cd.set_adapter('coupling')
outputs_cou = self.model_cd.base_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
self.model_cd.set_adapter('classify')
outputs_cla = self.model_cd.base_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
self.model_cd.set_adapter('identify')
outputs_ide = self.model_cd.base_model(
**model_inputs,
return_dict=True,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
logits_cou = outputs_cou.logits[:, -1, :] # torch.Size([8, 130528])
logits_ide = outputs_ide.logits[:, -1, :]
logits_cla = outputs_cla.logits[:, -1, :]
jsd_ide = self.jensen_shannon_divergence(logits_cou, logits_ide)
jsd_cla = self.jensen_shannon_divergence(logits_cou, logits_cla)
jsd_ide = torch.clamp(jsd_ide, min=0, max=1)
jsd_cla = torch.clamp(jsd_cla, min=0, max=1)
next_token_logits = []
for i in range(len(self.alter_pattern)):
cou,ide,cla,cd = logits_cou[i],logits_ide[i],logits_cla[i], logits_cd[i]
max_token = torch.argmax(cou)
if max_token == 4 or max_token == 12:
next_logit = cou
elif self.alter_pattern[i] == 'cla':
if self.model_args.AAC:
next_logit = cou + self.model_args.alpha * (cou + (1 - jsd_cla[i]) * cla - jsd_ide[i] * ide - cd)
else:
next_logit = cou + self.model_args.alpha * (cou + cla - ide - cd)
elif self.alter_pattern[i] == 'ide':
if self.model_args.AAC:
next_logit = cou + self.model_args.alpha * (cou + (1 - jsd_ide[i]) * ide - jsd_cla[i] * cla - cd)
else:
next_logit = cou + self.model_args.alpha * (cou + ide - cla - cd)
next_logit[cou < self.model_args.max_rate * torch.max(cou)] = -1000
next_logit[ide < self.model_args.max_rate * torch.max(ide)] = -1000
next_logit[cla < self.model_args.max_rate * torch.max(cla)] = -1000
next_token_logits.append(next_logit)
next_token_logits = torch.stack(next_token_logits,dim=0)
outputs = outputs_cou
else:
raise
return next_token_logits, outputs