gpo.py

import multiprocessing
import os
import random
import uuid
import warnings
from functools import partial

import torch
from accelerate import PartialState
from accelerate.logging import get_logger
from transformers import AutoModelForCausalLM, AutoTokenizer, set_seed
from transformers.integrations import is_deepspeed_zero3_enabled
from trl import ModelConfig, get_peft_config

from src.callbacks.generate_examples import GenerateExamplesCallback
from src.callbacks.training_parameters_callback import ParameterStatsCallback
from src.configs.additional.gpo_args import GPOScriptArguments
from src.configs.gpo_config import GroupedPOConfig
from src.trainers.gpo_trainer import GroupedPOTrainer
from src.utils.datasets import load_datasets, prepare_generative_row
from src.utils.logger import setup_logging
from src.utils.model_preparation import setup_model_and_tokenizer, unfreeze_modules_by_patterns
from src.utils.yaml_args_parser import H4ArgumentParser

logger = get_logger(__name__)

os.environ['WANDB_RUN_ID'] = str(random.randint(100000, 999999))

DATASET_PROCESSING_THREADS = min(multiprocessing.cpu_count() // 2, 16)


def main():
    parser = H4ArgumentParser((GPOScriptArguments, GroupedPOConfig, ModelConfig))
    args, gpo_config, model_config = parser.parse()

    setup_logging(logger, gpo_config)
    set_seed(gpo_config.seed)  # in case of new tokens added without initialize...

    os.environ["WANDB_PROJECT"] = args.project_name
    os.environ['CLEARML_PROJECT'] = args.project_name
    os.environ['WANDB_NAME'] = gpo_config.run_name.split("/")[-1]
    os.environ['CLEARML_TASK'] = gpo_config.run_name.split("/")[-1]

    ################
    # Model & Tokenizer
    ################
    tokenizer = AutoTokenizer.from_pretrained(model_config.model_name_or_path)
    model = AutoModelForCausalLM.from_pretrained(
        model_config.model_name_or_path,
        torch_dtype=torch.bfloat16 if gpo_config.bf16 else torch.float16,
        attn_implementation=model_config.attn_implementation
    )

    setup_model_and_tokenizer(args, model, tokenizer)

    if model_config.use_peft:
        for n, p in model.named_parameters():
            p.requires_grad = False
        if model_config.lora_task_type != "CAUSAL_LM":
            warnings.warn(
                "You are using a `task_type` that is different than `CAUSAL_LM` for PEFT. This will lead to silent bugs"
                " Make sure to pass --lora_task_type CAUSAL_LM when using this script."
            )
        if args.unfreeze_layers_patterns:
            warnings.warn(
                "You can't use non-empty unfreeze_layers_patterns and peft together at this time, only peft config will be used"
            )
        peft_config = get_peft_config(model_config)
    else:
        if args.unfreeze_layers_patterns:
            unfreeze_modules_by_patterns(model, args.unfreeze_layers_patterns)
        peft_config = None

    if PartialState().is_main_process:
        logger.info(f'Tokenizer: {tokenizer}')
        logger.info(f'Model config: {model.config}')
        logger.info(f'Model: {model}')

    ################
    # Dataset
    ################
    ds = load_datasets(args.dataset, args.test_size, args.dataset_ratio)
    generate_dataset = ds['test']

    def apply_chat_templates(row):
        row["prompt"] = tokenizer.apply_chat_template(row["prompt"], tokenize=False)
        row["completions"] = [tokenizer.apply_chat_template(chosen, tokenize=False) for chosen in row["completions"]]
        return row

    with PartialState().main_process_first():
        ds = ds.map(
            apply_chat_templates,
            num_proc=DATASET_PROCESSING_THREADS,
            keep_in_memory=True,
            load_from_cache_file=True
        )
        generate_dataset = generate_dataset.map(
            partial(prepare_generative_row, tokenizer=tokenizer, max_length=gpo_config.max_prompt_length),
            num_proc=DATASET_PROCESSING_THREADS,
            keep_in_memory=True,
            load_from_cache_file=True
        )

    train_dataset = ds["train"]
    eval_dataset = ds["test"]

    if PartialState().is_main_process:
        logger.info('Example from train dataset:')
        logger.info(train_dataset[0])
        logger.info('Example from test dataset:')
        logger.info(eval_dataset[0])
        logger.info('Example from gen dataset:')
        logger.info(generate_dataset[0])

    generate_callback = GenerateExamplesCallback(
        preprocessed_dataset=generate_dataset,
        tokenizer=tokenizer,
        num_examples=args.num_gen_examples,
        is_deepspeed_zero3=is_deepspeed_zero3_enabled(),
        logger_backend=gpo_config.report_to[0]
    )

    PartialState().wait_for_everyone()

    ################
    # Training
    ################
    trainer = GroupedPOTrainer(
        model,
        args=gpo_config,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        processing_class=tokenizer,
        peft_config=peft_config,
        callbacks=[generate_callback, ParameterStatsCallback] if args.generate_eval_examples else [ParameterStatsCallback]
    )

    # train and save the model
    trainer.train()

    if trainer.is_fsdp_enabled:
        trainer.accelerator.state.fsdp_plugin.set_state_dict_type("FULL_STATE_DICT")

    trainer.save_model(gpo_config.output_dir)


if __name__ == '__main__':
    main()