File size: 9,987 Bytes
a12c07f |
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 |
import argparse
import os
# os.environ["CUDA_VISIBLE_DEVICES"] = "0"
import json
import random
import re
import torch
import numpy as np
from tqdm import tqdm
import shortuuid
import sys
from transformers import AutoModel, AutoTokenizer, AutoProcessor, AutoModelForSeq2SeqLM
from torch.utils.data import Dataset, DataLoader
from PIL import Image
import math
from .gpt4v import TaskSpec, ParsedAnswer, Question
from .exceptions import GPTOutputParseException, GPTMaxTriesExceededException
import threading
from typing import List, Tuple, Union
from loguru import logger
from copy import deepcopy
import time
import os
seed = 42
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
class MiniCPMModel(object):
def __init__(self, task:TaskSpec,
model:str = "openbmb/MiniCPM-V-2_6-int4"):
self.task:TaskSpec = task
self.model = self.get_model(model)
self.tokenizer = AutoTokenizer.from_pretrained('openbmb/MiniCPM-V-2_6-int4', trust_remote_code=True)
def get_model(self, model):
# Load the open-source model in
if model == 'openbmb/MiniCPM-V-2_6-int4':
model_weights = AutoModel.from_pretrained(model, trust_remote_code=True,
attn_implementation='flash_attention_2', torch_dtype=torch.float16) # sdpa or flash_attention_2, no eager
model_weights = model_weights.eval()
return model_weights
else:
raise ValueError(f"Such model {model} does not exist!")
def ask(self, payload:dict, n_choices=1, temperature=0.7) -> Tuple[List[dict], List[dict]]:
"""
args:
payload: json dictionary, prepared by `prepare_payload`
"""
def minicpm_thread(self, idx, payload, results, temperature):
# creation of payload
mod_payload = deepcopy(payload)
messages = payload['messages']
max_tokens = payload['max_tokens']
try:
# Preparation for inference
output_text = self.model.chat(
image=None,
msgs=messages,
tokenizer=self.tokenizer
)
except Exception as e:
raise e
# print('outputs: ', output_text)
message = {'content' : output_text}
results[idx] = {"metadata": output_text, "message": message}
return
assert n_choices >= 1
results = [None] * n_choices
if n_choices > 1:
minicpm_jobs = [threading.Thread(target=minicpm_thread,
args=(self, idx, payload, results, temperature))
for idx in range(n_choices)]
for job in minicpm_jobs:
job.start()
for job in minicpm_jobs:
job.join()
else:
minicpm_thread(self, 0, payload, results, temperature)
messages:List[dict] = [ res["message"] for res in results]
metadata:List[dict] = [ res["metadata"] for res in results]
return messages, metadata
@staticmethod
def prepare_payload(question:Question,
max_tokens=1000,
verbose:bool=False,
prepend:Union[dict, None]=None,
**kwargs
) -> dict:
image_dic = None
text = ''
dic_list = question.get_json()
img_list = []
for dic in question.get_json():
# The case of text
if dic['type'] == 'text':
text += dic['text']
# The case of vision input
elif dic['type'] == 'image_url':
img_list.append(dic['image'])
if len(img_list) == 0:
img_list.append(Image.new('RGB', (512, 512), color = (255, 255, 255)))
content = [image for image in img_list]
content.append(text)
payload = {
"messages": [
{
'role': 'user',
"content":content,
},
],
"max_tokens": max_tokens,
}
return payload
def rough_guess(self, question:Question, max_tokens=1000,
max_tries=1, query_id:int=0,
verbose=False, temperature=1,
**kwargs):
p = self.prepare_payload(question, max_tokens = max_tokens, verbose=verbose, prepend=None,
model=self.model)
ok = False
reattempt = 0
while not ok:
response, meta_data = self.ask(p, temperature=temperature)
response = response[0]
# logger.info(f'response: {response}')
try:
parsed_response = self.task.answer_type.parser(response["content"])
except GPTOutputParseException as e:
# logger.warning(f"The response was not parseable:\n\n{response}\n\nBecause\n\n{e}")
# logger.warning(f"The response from LM was not parseable.")
pass
# if not os.path.exists('errors/'):
# # Create the directory if it doesn't exist
# os.makedirs('errors/')
# error_saved = f'errors/{time.strftime("%Y-%m-%d-%H-%M-%S")}.json'
# with open(error_saved, "w") as f:
# f.write(p_ans.code)
# logger.warning(f"The following was not parseable. Saved in {error_saved}.")
reattempt += 1
if reattempt > max_tries:
logger.error(f"max tries ({max_tries}) exceeded.")
raise GPTMaxTriesExceededException
logger.warning(f"Attempt failed: Reattempt #{reattempt} querying LLM")
continue
ok = True
return parsed_response, response, meta_data, p
def all_task_rough_guess(self, task, question:Question, max_tokens=1000,
max_tries=1, query_id:int=0,
verbose=False, temperature=1,
**kwargs):
p = self.prepare_payload(question, max_tokens = max_tokens, verbose=verbose, prepend=None,
model=self.model)
ok = False
reattempt = 0
while not ok:
response, meta_data = self.ask(p, temperature=temperature)
response = response[0]
# logger.info(f'response: {response}')
try:
parsed_response = task.answer_type.parser(response["content"])
except GPTOutputParseException as e:
# logger.warning(f"The following was not parseable:\n\n{response}\n\nBecause\n\n{e}")
# logger.warning(f"The response from LM was not parseable.")
pass
# if not os.path.exists('errors/'):
# # Create the directory if it doesn't exist
# os.makedirs('errors/')
# error_saved = f'errors/{time.strftime("%Y-%m-%d-%H-%M-%S")}.json'
# with open(error_saved, "w") as f:
# f.write(p_ans.code)
# logger.warning(f"The following was not parseable. Saved in {error_saved}.")
reattempt += 1
if reattempt > max_tries:
logger.error(f"max tries ({max_tries}) exceeded.")
raise GPTMaxTriesExceededException
logger.warning(f"Attempt failed: reattempt #{reattempt} querying LLM")
continue
ok = True
return parsed_response, response, meta_data, p
def many_rough_guesses(self, num_threads:int,
question:Question, max_tokens=1000,
verbose=False, max_tries=1, temperature=1
) -> List[Tuple[ParsedAnswer, str, dict, dict]]:
"""
Args:
num_threads : number of independent threads.
all other arguments are same as those of `rough_guess()`
Returns
List of elements, each element is a tuple following the
return signature of `rough_guess()`
"""
p = self.prepare_payload(question, max_tokens = max_tokens, verbose=verbose, prepend=None,
model=self.model)
# TODO
n_choices = num_threads
# TODO: wrap in robust-ask method, repeatedly asks until parseable output.
ok = False
reattempt = 0
while not ok:
response, meta_data = self.ask(p, n_choices=n_choices, temperature=temperature)
try:
parsed_response = [self.task.answer_type.parser(r["content"]) for r in response]
except GPTOutputParseException as e:
# logger.warning(f"The following was not parseable:\n\n{response}\n\nBecause\n\n{e}")
# logger.warning(f"The response from LM was not parseable.")
pass
# TODO provide the parse error message into GPT for the next round to be parsable
reattempt += 1
if reattempt > max_tries:
logger.error(f"max tries ({max_tries}) exceeded.")
raise GPTMaxTriesExceededException
logger.warning(f"Attempt failed: Reattempt #{reattempt} querying LLM")
continue
ok = True
return parsed_response, response, meta_data, p
def run_once(self, question:Question, max_tokens=1000, temperature=1, **kwargs):
q = self.task.first_question(question)
p_ans, ans, meta, p = self.rough_guess(q, max_tokens=max_tokens, temperature=temperature, **kwargs)
return p_ans, ans, meta, p
|