File size: 21,077 Bytes
349aa7a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 | import os
import re
import pathlib
from argparse import ArgumentParser
from typing import List, Dict, Optional
from dataclasses import dataclass, field
import torch
from torch import nn
import torch.nn.functional as F
from torch.optim import AdamW
from torch.utils.data import DataLoader, Dataset
from transformers import get_cosine_schedule_with_warmup, AutoTokenizer
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
AutoModelForMaskedLM,
AutoProcessor,
)
from datasets import load_dataset, DatasetDict
from peft import get_peft_model, LoraConfig, prepare_model_for_kbit_training
from transformers import BitsAndBytesConfig
import pytorch_lightning as pl
from pytorch_lightning.callbacks import ModelCheckpoint, LearningRateMonitor
from pytorch_lightning.loggers import WandbLogger
from trl import GRPOConfig, GRPOTrainer, ModelConfig, ScriptArguments, TrlParser, get_peft_config
# Import BLIP2 modules
from model.blip2_stage2 import Blip2Stage2
from blip2_dna_module import Blip2DNAModule
from blip2_grpo_trainer import Blip2GRPOTrainer
from bioreason.trainer import DNALLMGRPOConfig
# Custom TrainerCallback to override the saving mechanism
from transformers import TrainerCallback, TrainerState, TrainerControl
from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
class SaveWithPyTorchCallback(TrainerCallback):
"""Custom callback to save models with PyTorch's native save mechanism instead of safetensors"""
def on_save(self, args, state, control, **kwargs):
# Get the checkpoint folder
checkpoint_folder = os.path.join(
args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}"
)
os.makedirs(checkpoint_folder, exist_ok=True)
# Save with PyTorch instead of safetensors
checkpoint_path = os.path.join(checkpoint_folder, "pytorch_model.bin")
model = kwargs.get("model")
# Get model unwrapped from accelerator etc.
unwrapped_model = model.module if hasattr(model, "module") else model
# Save using PyTorch directly
torch.save(unwrapped_model.state_dict(), checkpoint_path)
# For BLIP2, save the config from the LLM component
if hasattr(unwrapped_model, "blip2") and hasattr(unwrapped_model.blip2, "llm_model"):
if hasattr(unwrapped_model.blip2.llm_model, "config"):
unwrapped_model.blip2.llm_model.config.save_pretrained(checkpoint_folder)
elif hasattr(unwrapped_model.blip2.llm_model, "base_model") and hasattr(unwrapped_model.blip2.llm_model.base_model, "config"):
unwrapped_model.blip2.llm_model.base_model.config.save_pretrained(checkpoint_folder)
# Print info about what's being saved
print(f"Saved model checkpoint to {checkpoint_folder}")
lora_params = [k for k in unwrapped_model.state_dict().keys() if "lora" in k]
print(f"Checkpoint contains {len(lora_params)} LoRA parameters")
# Signal that we've saved
control.should_save = False
return control
def extract_xml_answer(text: str) -> str:
answer = text.split("</think>")[-1]
return answer.strip()
def extract_hash_answer(text: str) -> str | None:
if "####" not in text:
return None
return text.split("####")[1].strip()
def get_kegg_questions() -> Dataset:
data = load_dataset('wanglab/kegg', 'default') # type: ignore
example_dna_sequences = ["ATCTACATGCAT", "CAGCAGCTACAG", "CATCACATCGACATCGAC"]
num_dna_sequences = 2
data = data.map(lambda x: { # type: ignore
'prompt': [
{
'role': 'user',
'content': [
*({'type': 'dna', 'text': None} for _ in range(num_dna_sequences)),
{'type': 'text', 'text': x['question']},
],
},
],
'dna_sequences': [x['reference_sequence'], x['variant_sequence']],
'answer': x['answer'],
}) # type: ignore
return data
def get_gsm8k_questions(question_prompt: str) -> Dataset:
data = load_dataset('openai/gsm8k', 'main') # type: ignore
example_dna_sequences = ["ATCTACATGCAT", "CAGCAGCTACAG", "CATCACATCGACATCGAC"]
data = data.map(lambda x: { # type: ignore
'prompt': [
{
'role': 'user',
'content': [
*({'type': 'dna', 'text': None} for _ in range(len(example_dna_sequences))),
{'type': 'text', 'text': 'Give me a short introduction to large language model.'}
]
},
],
'dna_sequences': [dna for dna in example_dna_sequences],
'answer': extract_hash_answer(x['answer']),
}) # type: ignore
return data # type: ignore
# Reward functions
def format_correct_reward_func(completions, **kwargs) -> list[float]:
"""
奖励函数:检查格式是否正确
要求:包含 <think>...</think> 和 <answer>...</answer> 标签
"""
responses = [completion[0]["content"] for completion in completions]
rewards = []
for response in responses:
score = 0.0
# 检查是否有think标签
if "<think>" in response and "</think>" in response:
score += 0.5
# 检查是否有answer标签
if "<answer>" in response and "</answer>" in response:
score += 0.5
# 检查标签的顺序是否正确
think_start = response.find("<think>")
think_end = response.find("</think>")
answer_start = response.find("<answer>")
answer_end = response.find("</answer>")
if (think_start != -1 and think_end != -1 and
answer_start != -1 and answer_end != -1 and
think_start < think_end < answer_start < answer_end):
score += 0.5 # 格式完全正确的额外奖励
rewards.append(score)
return rewards
def accuracy_reward_func(prompts, completions, answer, **kwargs) -> list[float]:
"""
奖励函数:检查答案准确率
"""
responses = [completion[0]['content'] for completion in completions]
rewards = []
for i, response in enumerate(responses):
# 提取answer标签中的内容
answer_match = re.search(r"<answer>(.*?)</answer>", response, re.DOTALL)
if answer_match:
extracted_answer = answer_match.group(1).strip()
else:
extracted_answer = response.strip()
# 获取正确答案
if isinstance(answer, list) and len(answer) > i:
correct_answer = str(answer[i]).strip()
elif isinstance(answer, list) and len(answer) > 0:
correct_answer = str(answer[0]).strip()
else:
correct_answer = str(answer).strip()
# 计算准确率奖励
if correct_answer.lower() in extracted_answer.lower():
rewards.append(1.0) # 完全匹配
elif any(word in extracted_answer.lower() for word in correct_answer.lower().split()):
rewards.append(0.5) # 部分匹配
else:
rewards.append(0.0) # 不匹配
return rewards
def repetition_penalty_reward_func(completions, **kwargs) -> list[float]:
"""
奖励函数:检查重复率(越低越好)
计算文本中重复词汇的比例,重复率越低奖励越高
"""
responses = [completion[0]["content"] for completion in completions]
rewards = []
for response in responses:
# 提取answer部分的文本
answer_match = re.search(r"<answer>(.*?)</answer>", response, re.DOTALL)
if answer_match:
text_to_analyze = answer_match.group(1).strip()
else:
text_to_analyze = response.strip()
# 分词并计算重复率
words = text_to_analyze.lower().split()
if len(words) == 0:
rewards.append(0.0)
continue
# 计算词汇重复率
unique_words = set(words)
repetition_rate = 1.0 - (len(unique_words) / len(words))
# 计算句子重复率
sentences = [s.strip() for s in text_to_analyze.split('.') if s.strip()]
if len(sentences) > 1:
unique_sentences = set(sentences)
sentence_repetition_rate = 1.0 - (len(unique_sentences) / len(sentences))
else:
sentence_repetition_rate = 0.0
# 综合重复率
overall_repetition = (repetition_rate + sentence_repetition_rate) / 2
# 重复率越低,奖励越高
reward = max(0.0, 1.0 - overall_repetition * 2) # 乘以2让惩罚更明显
rewards.append(reward)
return rewards
def combined_reward_func(prompts, completions, answer,
format_weight=0.3, accuracy_weight=0.5, repetition_weight=0.2,
**kwargs) -> list[float]:
"""
组合奖励函数:格式+准确率+重复率的加权组合
"""
format_rewards = format_correct_reward_func(completions, **kwargs)
accuracy_rewards = accuracy_reward_func(prompts, completions, answer, **kwargs)
repetition_rewards = repetition_penalty_reward_func(completions, **kwargs)
# 确保权重总和为1
total_weight = format_weight + accuracy_weight + repetition_weight
if total_weight != 1.0:
format_weight /= total_weight
accuracy_weight /= total_weight
repetition_weight /= total_weight
print(f"Normalized weights - Format: {format_weight:.3f}, Accuracy: {accuracy_weight:.3f}, Repetition: {repetition_weight:.3f}")
combined_rewards = []
for f_reward, a_reward, r_reward in zip(format_rewards, accuracy_rewards, repetition_rewards):
combined = (format_weight * f_reward +
accuracy_weight * a_reward +
repetition_weight * r_reward)
combined_rewards.append(combined)
return combined_rewards
# 保留一些原有的奖励函数作为备选
def less_than_4_reward_func(completions, **kwargs) -> list[float]:
responses = [completion[0]['content'] for completion in completions]
extracted_responses = [extract_xml_answer(r) for r in responses]
return [0.5 if len(r.split(' ')) <= 4 else 0.0 for r in extracted_responses]
def strict_format_reward_func(completions, **kwargs) -> list[float]:
"""Reward function that checks if the completion has a specific format."""
pattern = r"^<think>\n.*?\n</think>\n.*?\n$"
responses = [completion[0]["content"] for completion in completions]
matches = [re.match(pattern, r) for r in responses]
return [0.5 if match else 0.0 for match in matches]
def xmlcount_reward_func(completions, **kwargs) -> list[float]:
contents = [completion[0]["content"] for completion in completions]
return [count_xml(c) for c in contents]
def count_xml(text) -> float:
count = 0.0
if text.count("<think>\n") == 1:
count += 0.125
if text.count("\n</think>\n") == 1:
count += 0.125
return count
@dataclass
class Blip2ModelConfig(ModelConfig):
# BLIP2 specific configuration
model_name_or_path: str = field(default="blip2-model", metadata={"help": "Model checkpoint for weights initialization."})
# BLIP2 Architecture parameters
bert_name: str = field(default="/path/to/bert", metadata={"help": "BERT model for Q-former"})
num_query_token: int = field(default=32, metadata={"help": "Number of query tokens"})
cross_attention_freq: int = field(default=2, metadata={"help": "Cross attention frequency"})
plm_model: str = field(default="facebook/esm2_t30_150M_UR50D", metadata={"help": "Protein language model"})
plm_tune: str = field(default="freeze", metadata={"help": "PLM tuning strategy"})
llm_name: str = field(default="facebook/galactica-1.3b", metadata={"help": "Language model name"})
llm_tune: str = field(default="lora", metadata={"help": "LLM tuning strategy"})
qformer_tune: str = field(default="train", metadata={"help": "Q-former tuning strategy"})
peft_dir: str = field(default="", metadata={"help": "PEFT directory"})
# LoRA parameters
lora_r: int = field(default=8, metadata={"help": "LoRA rank"})
lora_alpha: int = field(default=16, metadata={"help": "LoRA alpha"})
lora_dropout: float = field(default=0.1, metadata={"help": "LoRA dropout"})
# Training parameters
enbale_gradient_checkpointing: bool = field(default=False, metadata={"help": "Enable gradient checkpointing"})
enable_flash: bool = field(default=False, metadata={"help": "Enable flash attention"})
# Other parameters
cache_dir: str = field(default=None, metadata={"help": "Path to model cache directory."})
sft_checkpoint: str = field(default=None, metadata={"help": "Path to the checkpoint for SFT."})
freeze_dna_modules: bool = field(default=False, metadata={"help": "Freeze DNA/protein modules"})
@dataclass
class GRPOScriptArguments(ScriptArguments):
"""
Script arguments for the GRPO training script with BLIP2.
"""
dataset_name: str = field(default="wanglab/kegg", metadata={"help": "Dataset name with default."})
data_file_paths: str = field(
default=None,
metadata={"help": "Paths to data files, separated by ':'"},
)
arrow_cache_dir: str = field(
default=None,
metadata={"help": "Path to arrow cache directory"},
)
val_split_ratio: float = field(
default=0.0,
metadata={"help": "Ratio of validation split, default 0.0"},
)
reward_funcs: list[str] = field(
# 选项1:使用组合奖励函数(推荐)
default_factory=lambda: ["combined"],
# 选项2:使用分离的三个奖励函数
# default_factory=lambda: ["format_correct", "accuracy", "repetition_penalty"],
# 选项3:自定义组合
# default_factory=lambda: ["format_correct", "accuracy", "repetition_penalty", "xmlcount"],
metadata={"help": "List of reward functions. Available: 'combined', 'format_correct', 'accuracy', 'repetition_penalty', 'xmlcount', 'strict_format', 'less_than_4'"},
)
# 奖励函数权重配置
format_weight: float = field(
default=0.3,
metadata={"help": "Weight for format correctness reward (used in combined reward)"}
)
accuracy_weight: float = field(
default=0.5,
metadata={"help": "Weight for accuracy reward (used in combined reward)"}
)
repetition_weight: float = field(
default=0.2,
metadata={"help": "Weight for repetition penalty reward (used in combined reward)"}
)
reward_funcs_registry = {
# 新的三合一奖励函数
"combined": combined_reward_func, # 格式+准确率+重复率组合
# 分离的奖励函数
"format_correct": format_correct_reward_func, # 格式正确性
"accuracy": accuracy_reward_func, # 准确率
"repetition_penalty": repetition_penalty_reward_func, # 重复率惩罚
# 原有的奖励函数(保留作为备选)
"xmlcount": xmlcount_reward_func,
"strict_format": strict_format_reward_func,
"less_than_4": less_than_4_reward_func,
}
def get_vlm_module(model_name_or_path):
# Always use BLIP2 module for this implementation
return Blip2DNAModule
def create_blip2_args_from_config(model_args):
"""Create BLIP2 args from model config"""
# Convert model config to the format expected by BLIP2
blip2_args = {
'bert_name': model_args.bert_name,
'num_query_token': model_args.num_query_token,
'cross_attention_freq': model_args.cross_attention_freq,
'plm_model': model_args.plm_model,
'plm_tune': model_args.plm_tune,
'llm_name': model_args.llm_name,
'llm_tune': model_args.llm_tune,
'qformer_tune': model_args.qformer_tune,
'peft_dir': model_args.peft_dir,
'lora_r': model_args.lora_r,
'lora_alpha': model_args.lora_alpha,
'lora_dropout': model_args.lora_dropout,
'enbale_gradient_checkpointing': model_args.enbale_gradient_checkpointing,
'enable_flash': model_args.enable_flash,
}
return blip2_args
def _prep_for_training(model, training_args):
"""
Prepare BLIP2 model for training with LoRA.
"""
# The BLIP2 model should handle its own LoRA setup
# This is mainly for any additional preparation needed
target_modules = ["q_proj", "k_proj", "v_proj", "out_proj", "fc1", "fc2"]
lora_config = LoraConfig(
r=training_args.lora_r,
lora_alpha=training_args.lora_alpha,
lora_dropout=training_args.lora_dropout,
target_modules=target_modules,
init_lora_weights="gaussian",
bias="none",
task_type="CAUSAL_LM",
)
return lora_config
def main(script_args, training_args, model_args):
print(training_args.output_dir)
torch.cuda.empty_cache()
torch.set_float32_matmul_precision("medium")
# Create BLIP2 model
blip2_args = create_blip2_args_from_config(model_args)
model = Blip2Stage2(blip2_args)
# Load checkpoint if specified
if model_args.sft_checkpoint is not None:
print(f"Loading SFT checkpoint from {model_args.sft_checkpoint}")
if os.path.isdir(model_args.sft_checkpoint):
# Load Lightning checkpoint
checkpoint = torch.load(os.path.join(model_args.sft_checkpoint, "last.ckpt"), map_location='cpu')
model.load_state_dict(checkpoint['state_dict'], strict=False)
print("Loaded Lightning checkpoint")
else:
# Load PyTorch state dict
checkpoint = torch.load(model_args.sft_checkpoint, map_location='cpu')
if "state_dict" in checkpoint:
state_dict = checkpoint["state_dict"]
else:
state_dict = checkpoint
# Remove module prefix if present
state_dict = {k.replace("module.", ""): v for k, v in state_dict.items()}
result = model.load_state_dict(state_dict, strict=False)
print(f"Loaded checkpoint with {len(result.missing_keys)} missing keys and {len(result.unexpected_keys)} unexpected keys")
# Get reward functions with weights
reward_funcs = []
for func_name in script_args.reward_funcs:
if func_name == "combined":
# 为组合奖励函数传递权重参数
def weighted_combined_reward(prompts, completions, answer, **kwargs):
return combined_reward_func(
prompts, completions, answer,
format_weight=script_args.format_weight,
accuracy_weight=script_args.accuracy_weight,
repetition_weight=script_args.repetition_weight,
**kwargs
)
reward_funcs.append(weighted_combined_reward)
else:
reward_funcs.append(reward_funcs_registry[func_name])
print("reward_funcs:", [func.__name__ if hasattr(func, '__name__') else 'weighted_combined_reward' for func in reward_funcs])
print(f"Reward weights - Format: {script_args.format_weight}, Accuracy: {script_args.accuracy_weight}, Repetition: {script_args.repetition_weight}")
vlm_module_cls = get_vlm_module(model_args.model_name_or_path)
print("using vlm module:", vlm_module_cls.__name__)
question_prompt = vlm_module_cls.get_question_template()
# Load dataset
dataset = get_kegg_questions()
print(dataset)
# Custom callback to handle saving with PyTorch's native mechanism
custom_save_callback = SaveWithPyTorchCallback()
# Initialize the BLIP2 GRPO trainer
trainer = Blip2GRPOTrainer(
model=model,
reward_funcs=reward_funcs,
args=training_args,
dna_module=vlm_module_cls(),
train_dataset=dataset['train'],
eval_dataset=dataset['val'] if training_args.eval_strategy != "no" else None,
peft_config=get_peft_config(model_args),
attn_implementation=getattr(model_args, 'attn_implementation', 'flash_attention_2'),
torch_dtype=getattr(model_args, 'torch_dtype', 'bfloat16'),
callbacks=[custom_save_callback],
)
# Set the trainer to save in PyTorch format instead of safetensors
training_args.save_safetensors = False
# Train the model
trainer.train()
if __name__ == "__main__":
print(f"CUDA_VISIBLE_DEVICES: {os.environ.get('CUDA_VISIBLE_DEVICES')}")
parser = TrlParser((GRPOScriptArguments, DNALLMGRPOConfig, Blip2ModelConfig))
script_args, training_args, model_args = parser.parse_args_and_config()
# Ensure we use PyTorch's save mechanism instead of safetensors
training_args.save_safetensors = False
main(script_args, training_args, model_args) |