-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
175 lines (158 loc) · 6.52 KB
/
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
import functools
import logging
import random
import string
import time
from transformers import RobertaTokenizer, T5ForConditionalGeneration
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, AutoModelForCausalLM
import torch
device = "cuda" if torch.cuda.is_available() else "cpu"
@functools.lru_cache()
def get_codegen():
"""Return the codegen tokenizer and model"""
logger = logging.getLogger("pilot.model")
modelname = 'Salesforce/codegen-2B-mono'
tokenizer = AutoTokenizer.from_pretrained(modelname)
model = AutoModelForCausalLM.from_pretrained(modelname,
#load_in_8bit=True, # <- needs bitsandbytes and accelerate, and libcudart
torch_dtype=torch.float16,
low_cpu_mem_usage=True,
#device_map="auto"
).to(device)
logger.info("Loaded %s, footprint %s", modelname, model.get_memory_footprint())
return tokenizer, model
@functools.lru_cache()
def get_codegen2():
"""Return the codegen2 tokenizer and model"""
logger = logging.getLogger("pilot.model")
modelname = 'Salesforce/codegen2-1B'
tokenizer = AutoTokenizer.from_pretrained(modelname)
model = AutoModelForCausalLM.from_pretrained(modelname, trust_remote_code=True, revision="main").to(device)
logger.info("Loaded %s, footprint %s", modelname, model.get_memory_footprint())
return tokenizer, model
@functools.lru_cache()
def get_codet5():
"""Return the codet5-base tokenizer and model"""
logger = logging.getLogger("pilot.model")
modelname = 'Salesforce/codet5-base'
tokenizer = RobertaTokenizer.from_pretrained(modelname)
model: T5ForConditionalGeneration = T5ForConditionalGeneration.from_pretrained(modelname).to(device)
logger.info("Loaded %s, footprint %s", modelname, model.get_memory_footprint())
return tokenizer, model
@functools.lru_cache()
def get_codet5p():
"""Return the codet5+ tokenizer and model"""
logger = logging.getLogger("pilot.model")
modelname = 'Salesforce/codet5p-770m'
tokenizer = AutoTokenizer.from_pretrained(modelname)
model = AutoModelForSeq2SeqLM.from_pretrained(modelname,
#torch_dtype=torch.float16,
#low_cpu_mem_usage=True,
trust_remote_code=True).to(device)
logger.info("Loaded %s, footprint %s", modelname, model.get_memory_footprint())
return tokenizer, model
def dummy_generate(data):
"""Run dummy model: for fast testing"""
text = data['prompt']
completion = "<this is a dummy output>"
response = {
"id": 'cmpl-' + ''.join(random.choice(string.ascii_letters + string.digits) for _ in range(29)),
"model": data["model"],
"object": "text_completion",
"created": int(time.time()),
"choices": [{
'text': completion,
'index': 0,
'finish_reason': 'stop',
'logprobs': None
}],
"usage": {
"completion_tokens": len(completion),
"prompt_tokens": len(text),
"total_tokens": len(text)+len(completion),
}
}
return response
def generate(data):
"""Run model"""
logger = logging.getLogger("pilot.run")
logger.debug(data)
# Load model and tokenizer
if data['model'] == 'codegen':
tokenizer, model = get_codegen()
elif data['model'] == 'codegen2':
tokenizer, model = get_codegen2()
elif data['model'] == 'codet5':
tokenizer, model = get_codet5()
elif data['model'] == 'codet5p':
tokenizer, model = get_codet5p()
elif data['model'] == 'dummy':
return dummy_generate(data)
else:
raise NotImplementedError(f"Model {data['model']} unrecognized")
# get the prompt and prepare parameters
text = data['prompt']
logger.debug(text)
encoding = tokenizer(text, return_tensors="pt").to(device)
logger.debug(encoding)
attention_mask = encoding.get("attention_mask") # TODO ensure not None
kwargs = {}
if data['model'] == 'codet5p':
kwargs.update(encoding)
kwargs['decoder_input_ids'] = kwargs['input_ids'].clone()
else:
kwargs['inputs'] = encoding.input_ids
if data['suffix']:
# TODO handle data['suffix'] - need client plugin to support
raise NotImplementedError(f"suffix is not supported yet")
# update other input parameters
if data['temperature'] is not None:
kwargs['temperature'] = data['temperature']
if kwargs['temperature'] == 0.0:
kwargs['temperature'] = 1.0
kwargs['top_k'] = 1
else:
kwargs['temperature'] = 0.2
if data['max_tokens'] is not None:
kwargs['max_new_tokens'] = data['max_tokens']
kwargs['top_p'] = data['top_p'] or 1.0
if data['presence_penalty'] is not None:
kwargs['diversity_penalty'] = data['presence_penalty']
if data['frequency_penalty'] is not None:
kwargs['repetition_penalty'] = data['frequency_penalty']
else:
kwargs['repetition_penalty'] = 1.0
kwargs['do_sample'] = True
kwargs['num_return_sequences'] = 1 # what if more than 1?
kwargs['attention_mask'] = attention_mask
kwargs['pad_token_id'] = tokenizer.eos_token_id
# run model and decode generated content
logger.debug(kwargs)
generated = model.generate(**kwargs)
logger.debug(generated)
prompt_tokens = len(encoding["input_ids"][0])
generated = generated[0]
if data['model'] in ['codegen', 'codegen2', 'codet5p']:
# these models will repeat the prompt
generated = generated[prompt_tokens:]
completion = tokenizer.decode(generated, skip_special_tokens=True)
completion_tokens = len(generated)
# prepare response
response = {
"id": 'cmpl-' + ''.join(random.choice(string.ascii_letters + string.digits) for _ in range(29)),
"model": data["model"],
"object": "text_completion",
"created": int(time.time()),
"choices": [{
'text': completion,
'index': 0,
'finish_reason': 'length' if completion_tokens==data['max_tokens'] else 'stop',
'logprobs':None
}],
"usage": {
"completion_tokens": completion_tokens,
"prompt_tokens": prompt_tokens,
"total_tokens": prompt_tokens+completion_tokens,
}
}
return response