aurora

1
# Inspired by: https://github.com/lvwerra/trl/blob/main/examples/research_projects/stack_llama/scripts/rl_training.py
2

3
import math
4
from trl import PPOConfig
5
from torch.optim import AdamW
6
from typing import TYPE_CHECKING, Optional, List
7
from transformers import DataCollatorWithPadding
8
from transformers.optimization import get_scheduler
9

10
from llmtuner.data import get_dataset, preprocess_dataset
11
from llmtuner.extras.callbacks import SavePeftModelCallback
12
from llmtuner.extras.ploting import plot_loss
13
from llmtuner.model import load_model_and_tokenizer
14
from llmtuner.train.utils import create_ref_model, create_reward_model
15
from llmtuner.train.ppo.trainer import CustomPPOTrainer
16

17
if TYPE_CHECKING:
18
    from transformers import Seq2SeqTrainingArguments, TrainerCallback
19
    from llmtuner.hparams import ModelArguments, DataArguments, FinetuningArguments, GeneratingArguments
20

21

22
def run_ppo(
23
    model_args: "ModelArguments",
24
    data_args: "DataArguments",
25
    training_args: "Seq2SeqTrainingArguments",
26
    finetuning_args: "FinetuningArguments",
27
    generating_args: "GeneratingArguments",
28
    callbacks: Optional[List["TrainerCallback"]] = None
29
):
30
    dataset = get_dataset(model_args, data_args)
31
    model, tokenizer = load_model_and_tokenizer(model_args, finetuning_args, training_args.do_train, add_valuehead=True)
32
    dataset = preprocess_dataset(dataset, tokenizer, data_args, training_args, stage="ppo")
33

34
    tokenizer.padding_side = "left" # use left-padding in generation while using right-padding in training
35
    data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
36

37
    # Create reference model and reward model
38
    ref_model = create_ref_model(model_args, finetuning_args, add_valuehead=True)
39
    reward_model = create_reward_model(model, model_args, finetuning_args)
40

41
    # Create ppo config
42
    backward_batch_size = training_args.per_device_train_batch_size * training_args.gradient_accumulation_steps
43
    ppo_config = PPOConfig(
44
        model_name=model_args.model_name_or_path,
45
        learning_rate=training_args.learning_rate,
46
        mini_batch_size=training_args.per_device_train_batch_size,
47
        batch_size=backward_batch_size * finetuning_args.ppo_buffer_size,
48
        gradient_accumulation_steps=training_args.gradient_accumulation_steps,
49
        ppo_epochs=finetuning_args.ppo_epochs,
50
        max_grad_norm=training_args.max_grad_norm,
51
        seed=training_args.seed,
52
        optimize_device_cache=True,
53
        target=finetuning_args.ppo_target,
54
        log_with=finetuning_args.ppo_logger,
55
        use_score_scaling=finetuning_args.ppo_score_norm,
56
        use_score_norm=finetuning_args.ppo_score_norm,
57
        whiten_rewards=finetuning_args.ppo_whiten_rewards,
58
        accelerator_kwargs={"step_scheduler_with_optimizer": False}
59
    )
60

61
    # Create optimizer and scheduler
62
    optimizer = AdamW(filter(lambda p: p.requires_grad, model.parameters()), lr=training_args.learning_rate)
63
    if training_args.max_steps > 0:
64
        num_training_steps = training_args.max_steps
65
    else:
66
        total_train_batch_size = backward_batch_size * finetuning_args.ppo_buffer_size * training_args.world_size
67
        num_training_steps = training_args.num_train_epochs * math.ceil(len(dataset) / total_train_batch_size)
68

69
    lr_scheduler = get_scheduler(
70
        training_args.lr_scheduler_type,
71
        optimizer=optimizer,
72
        num_warmup_steps=training_args.get_warmup_steps(num_training_steps),
73
        num_training_steps=num_training_steps
74
    )
75

76
    # Initialize our Trainer
77
    ppo_trainer = CustomPPOTrainer(
78
        model_args=model_args,
79
        training_args=training_args,
80
        finetuning_args=finetuning_args,
81
        generating_args=generating_args,
82
        callbacks=callbacks + [SavePeftModelCallback()],
83
        reward_model=reward_model,
84
        config=ppo_config,
85
        model=model,
86
        ref_model=ref_model,
87
        tokenizer=tokenizer,
88
        dataset=dataset,
89
        data_collator=data_collator,
90
        optimizer=optimizer,
91
        lr_scheduler=lr_scheduler
92
    )
93

94
    # Training
95
    if training_args.do_train:
96
        ppo_trainer.ppo_train(resume_from_checkpoint=training_args.resume_from_checkpoint)
97
        ppo_trainer.save_model()
98
        ppo_trainer.save_state() # must be called after save_model to have a folder
99
        if ppo_trainer.is_world_process_zero() and finetuning_args.plot_loss:
100
            plot_loss(training_args.output_dir, keys=["loss", "reward"])
101