-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcodellama_generate_manual.py
366 lines (315 loc) · 16.5 KB
/
codellama_generate_manual.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
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
import json
# os.environ["cuda_visible_devices"] = "1,2,3,4,5,6,7"
import os.path
from argparse import ArgumentParser
from pathlib import Path
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from datetime import datetime
from prompt_construction_infill import construct_prompt_infill
from util import HistoryCategory, bugs_fail_ground_test
from model_output_parser import (parse_output_codellama_34b_instruct,
parse_output_codellama_34b_python,
parse_output_codellama_70b_instruct,
parse_output_codellama_70b_python, parse_output_infill)
from prompt_construction import (construct_prompt_codellama_34b_instruct,
construct_prompt_codellama_34b_python,
construct_prompt_codellama_70b_instruct,
construct_prompt_codellama_70b_python)
def get_parser():
parser = ArgumentParser()
parser.add_argument('--prompt_style', type=str, choices=['instruct', 'infill'])
parser.add_argument('--subject_model_id', type=str)
parser.add_argument('--history_category', type=str)
parser.add_argument('--bugs_meta_data_file', type=str)
parser.add_argument('--bugs_description_file', type=str)
parser.add_argument('--history_data_path', type=Path)
parser.add_argument('--result_output_path', type=Path)
parser.add_argument('--has_nucleus_sampling', type=str)
parser.add_argument('--is_buggy_line_labeled', type=str)
parser.add_argument('--bug_id_list', nargs='+')
return parser
args = get_parser().parse_args()
def main():
# only 7B and 13B Code Llama and Code Llama - Instruct variants support infilling based on surrounding content.
gpu_num = torch.cuda.device_count()
print(f"The number of GPU available is: {gpu_num}\n")
# narrow the scope of subject projects
subject_projects = ["luigi", "youtube-dl", "keras", "thefuck"]
prompt_style = args.prompt_style
model_id = args.subject_model_id
history_category_flag = str(args.history_category)
bugs_meta_data_file = args.bugs_meta_data_file
bugs_description_file = args.bugs_description_file
history_data_path = args.history_data_path
result_output_path = args.result_output_path
has_nucleus_sampling = bool(int(args.has_nucleus_sampling))
is_buggy_line_labeled = bool(int(args.is_buggy_line_labeled))
bug_id_list = args.bug_id_list
print(f"function_level_history flag: {history_category_flag}")
print(f"Target bug_id_list is: {bug_id_list}")
# for model_id in subject_models:
print("==============================================================================================")
print(f"{model_id}, history category: {HistoryCategory(history_category_flag).name} start generation!")
model_generation(
prompt_style,
model_id,
subject_projects,
bugs_fail_ground_test,
history_category_flag,
bugs_meta_data_file,
bugs_description_file,
history_data_path,
result_output_path,
has_nucleus_sampling,
is_buggy_line_labeled,
bug_id_list
)
print(f"{model_id}, history category: {HistoryCategory(history_category_flag).name} finish generation!")
print("==============================================================================================")
def model_generation(prompt_style, model_id, subject_projects, bugs_fail_ground_test, history_category_flag, bugs_meta_data_file,
bugs_description_file, history_data_path, result_output_path,
has_nucleus_sampling, is_buggy_line_labeled, bug_id_list):
# prompt = f"{B_INST} {B_SYS}{SYSTEM_PROMPT}{E_SYS}{pre_prompt} {E_INST}"
# Prompt: https://www.promptingguide.ai/models/code-llama#configure-model-access
# prompt template for codellama-python: https://huggingface.co/TheBloke/CodeLlama-7B-Python-GGUF
# Performance comparison: https://huggingface.co/blog/codellama
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float16,
# device_map="balanced_low_0",
device_map="auto",
)
print(f"model device_map:\n{json.dumps(model.hf_device_map, indent=4)}\n")
tokenizer = AutoTokenizer.from_pretrained(model_id)
# quantized model
# model = AutoModelForCausalLM.from_pretrained(
# model_id,
# # torch_dtype=torch.float16,
# # device_map="balanced_low_0",
# device_map="auto",
# trust_remote_code=True,
# revision="main"
# )
# print(f"model device_map:\n{json.dumps(model.hf_device_map, indent=4)}\n")
# tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
# save the result
model_name = model_id.split('/')[1].replace('-', '_').lower()
target_path = os.path.join(result_output_path, f"{model_name}_{prompt_style}")
os.makedirs(target_path, exist_ok=True)
result_save_path = os.path.join(target_path, f"{HistoryCategory(history_category_flag).name}.json")
if os.path.exists(result_save_path):
os.remove(result_save_path)
bugs_meta_data: dict = json.load(open(bugs_meta_data_file, 'r'))
bugs_description_data: dict = json.load(open(bugs_description_file, 'r'))
result = {}
for bug_id, bug_value in bugs_meta_data.items():
if bug_id not in bug_id_list:
continue
# if bug_id != '1':
# break
# if bug_value['project_name'] not in subject_projects:
# continue
# if int(bug_id) < 60:
# continue
if bug_id in bugs_fail_ground_test:
continue
start_generation_time, start_generation_time_format = current_time()
print(f"Current time {start_generation_time_format}: Start generation\n")
code_length = bug_value['function']['function_before_token_count']
history_save_path = os.path.join(history_data_path, f"{bug_id}.json")
if not os.path.exists(history_save_path):
# id = 34, 39
# skip the bugs without history
print(f"bug {bug_id} has no history data.")
continue
history_meta_data: dict = json.load(open(history_save_path, 'r'))
bugs_description: dict = bugs_description_data[bug_id]
if prompt_style == "instruct":
if 'CodeLlama-34b-Instruct' in model_id or 'CodeLlama-13b-Instruct' in model_id or 'CodeLlama-7b-Instruct' in model_id:
prompt = construct_prompt_codellama_34b_instruct(
bug_value,
history_category_flag,
history_meta_data,
bugs_description,
is_buggy_line_labeled
)
input_tokens = tokenizer(
prompt,
return_tensors="pt",
add_special_tokens=True
).input_ids.to('cuda')
elif 'CodeLlama-34b-Python' in model_id:
prompt = construct_prompt_codellama_34b_python(bug_value, history_category_flag, history_meta_data,
bugs_description)
input_tokens = tokenizer(
prompt,
return_tensors="pt",
add_special_tokens=True
).input_ids.to('cuda')
elif 'CodeLlama-70b-Instruct' in model_id:
prompt = construct_prompt_codellama_70b_instruct(bug_value, history_category_flag, history_meta_data,
bugs_description)
input_tokens = tokenizer.apply_chat_template(
prompt,
return_tensors="pt"
).to("cuda")
elif 'CodeLlama-70b-Python' in model_id:
prompt = construct_prompt_codellama_70b_python(bug_value, history_category_flag, history_meta_data,
bugs_description)
input_tokens = tokenizer(
prompt,
return_tensors="pt",
add_special_tokens=True
).input_ids.to('cuda')
if int(code_length) < 100:
max_gen_len = 500
elif int(code_length) < 500:
max_gen_len = 1000
else:
max_gen_len = 1500
elif prompt_style == 'infill':
# infill
prompt, masked_buggy_code = construct_prompt_infill(bug_value, history_category_flag, history_meta_data,
bugs_description)
if history_category_flag == HistoryCategory.pure_infill.value:
prompt = masked_buggy_code
input_tokens = tokenizer(
prompt,
return_tensors="pt",
add_special_tokens=True
).input_ids.to('cuda')
max_gen_len = 100
num_tokens = input_tokens.shape[1]
print(f"{model_id}, history category: {HistoryCategory(history_category_flag).name}, prompt is:\n{prompt}")
return_sequences = 10
# torch.cuda.empty_cache()
# decoding strategy: https://michael-franke.github.io/npNLG/06-LSTMs/06d-decoding-GPT2.html
# 1. greedy search: do_sample = False and num_beams = 1
_, start_model_generate_time = current_time()
print(
f"Current time {start_model_generate_time}: Start model.generate() to generate fixed code for bug {bug_id}......")
try:
generation_output_greedy = model.generate(
input_ids=input_tokens,
max_new_tokens=max_gen_len,
)
except:
print(f"bug: {bug_id} meet error on greedy search!")
# torch.cuda.empty_cache()
print(f"current memory summary:\n{torch.cuda.memory_summary()}")
continue
generation_output_greedy_fixed_code = generation_output_greedy[0][input_tokens.shape[-1]:]
if prompt_style == "instruct":
result_greedy_original = tokenizer.decode(generation_output_greedy_fixed_code,
skip_special_tokens=True).strip()
elif prompt_style == 'infill':
result_greedy_original = tokenizer.decode(generation_output_greedy_fixed_code, skip_special_tokens=True)
finish_generation_time, finish_generation_time_format = current_time()
generation_period_minutes = time_period_minutes(finish_generation_time, start_generation_time)
print(f"Current time {finish_generation_time_format}: Finish generation\n")
print(
f"Input token size: {num_tokens}. Spend around {generation_period_minutes} minutes for bug {bug_id} on greedy inference!\n")
try:
if prompt_style == "instruct":
if 'CodeLlama-34b-Instruct' in model_id or 'CodeLlama-13b-Instruct' in model_id or 'CodeLlama-7b-Instruct' in model_id:
result_greedy_parsed = parse_output_codellama_34b_instruct(result_greedy_original)
elif 'CodeLlama-34b-Python' in model_id:
result_greedy_parsed = parse_output_codellama_34b_python(result_greedy_original)
elif 'CodeLlama-70b-Instruct' in model_id:
result_greedy_parsed = parse_output_codellama_70b_instruct(result_greedy_original)
elif 'CodeLlama-70b-Python' in model_id:
result_greedy_parsed = parse_output_codellama_70b_python(result_greedy_original)
elif prompt_style == 'infill':
result_greedy_parsed_infill_line = parse_output_infill(result_greedy_original)
result_greedy_parsed = masked_buggy_code.replace("<FILL_ME>", result_greedy_parsed_infill_line)
except:
print(f"bug {bug_id}'s greedy result cannot be parsed!")
result_greedy_parsed = ""
if has_nucleus_sampling:
# 2. top-p/nucleus sampling: do_sample = False and num_beams = 1
# torch.cuda.empty_cache()
try:
generation_output_nucleus = model.generate(
input_ids=input_tokens,
max_new_tokens=max_gen_len,
do_sample=True,
top_p=0.95,
top_k=50,
temperature=0.4,
num_return_sequences=return_sequences,
# repetition_penalty=1.2,
)
except:
print(f"bug: {bug_id} meet error on nucleus sampling!")
# torch.cuda.empty_cache()
print(f"current memory summary:\n{torch.cuda.memory_summary()}")
continue
result_nucleus_original = []
result_nucleus_parsed = []
for i, output in enumerate(generation_output_nucleus):
if prompt_style == "instruct":
decoded_result_i = tokenizer.decode(output[input_tokens.shape[-1]:],
skip_special_tokens=True).strip()
elif prompt_style == 'infill':
decoded_result_i = tokenizer.decode(output[input_tokens.shape[-1]:], skip_special_tokens=True)
result_nucleus_original.append(decoded_result_i)
try:
if prompt_style == "instruct":
if 'CodeLlama-34b-Instruct' in model_id or 'CodeLlama-13b-Instruct' in model_id or 'CodeLlama-7b-Instruct' in model_id:
decoded_result_i_parsed = parse_output_codellama_34b_instruct(decoded_result_i)
elif 'CodeLlama-34b-Python' in model_id:
decoded_result_i_parsed = parse_output_codellama_34b_python(decoded_result_i)
elif 'CodeLlama-70b-Instruct' in model_id:
decoded_result_i_parsed = parse_output_codellama_70b_instruct(decoded_result_i)
elif 'CodeLlama-70b-Python' in model_id:
decoded_result_i_parsed = parse_output_codellama_70b_python(decoded_result_i)
elif prompt_style == 'infill':
decoded_result_i_parsed_infill_line = parse_output_infill(decoded_result_i)
decoded_result_i_parsed = masked_buggy_code.replace("<FILL_ME>",
decoded_result_i_parsed_infill_line)
except:
print(f"bug {bug_id}'s {i}th nucleus result cannot be parsed!")
decoded_result_i_parsed = ""
result_nucleus_parsed.append(decoded_result_i_parsed)
else:
result_nucleus_parsed = ""
result_nucleus_original = ""
result[bug_id] = {
"id": bug_id,
"ground_fixed_code": bug_value['function']['function_after'],
"input": {
"has_buggy_code": 1,
"has_commit_msg": 1,
"has_function_level_history": 1 if history_category_flag == '1' else 0,
"has_file_level_history": 1 if history_category_flag == '2' else 0,
"buggy_code": bug_value['function']['function_before'],
"prompt": prompt
},
"output": {
"greedy_search": result_greedy_parsed,
"nucleus_sampling": result_nucleus_parsed
},
"output_original": {
"greedy_search": result_greedy_original,
"nucleus_sampling": result_nucleus_original
},
}
json.dump(result, open(result_save_path, 'w'), indent=2)
print(f"mode_id: {model_id}\n, history category: {HistoryCategory(history_category_flag).name}\n, "
f"finish the bug: {bug_id}!\n"
f"max_gen_len: {max_gen_len}\n\n"
f"decoding_greedy_search_original_output: \n{result_greedy_original}\n\n\n"
f"decoding_nucleus_sampling_original_output: \n{result_nucleus_original}"
)
print("===============================================================================")
torch.cuda.empty_cache()
def current_time():
current_time_ = datetime.now()
return current_time_, current_time_.strftime("%Y-%m-%d %H:%M:%S")
def time_period_minutes(end_time, start_time):
time_period = end_time - start_time
total_seconds = time_period.total_seconds()
return round(total_seconds / 60, 1)
if __name__ == "__main__":
main()