finetune.py

import os
import sys
from typing import List
from functools import partial
import fire
import torch
import transformers
from peft import PeftModel
from datasets import load_dataset, concatenate_datasets

from peft import (
    LoraConfig,
    get_peft_model,
    get_peft_model_state_dict,
    prepare_model_for_int8_training,
    set_peft_model_state_dict,
)
from transformers import AutoModelForCausalLM, AutoTokenizer

from prompter import Prompter

def train(
    # model/data params
    base_model: str = "bigscience/bloom-560m",
    data_path: str = "data/gsm8k/en/train.jsonl",
    output_dir: str = "output/bigscience/bloom-560m/xnli/en",
    prompt_lang: str = 'en',
    task_name: str = "xnli",
    # training hyperparams
    load_8bit: bool = False,
    batch_size: int = 32,
    micro_batch_size: int = 4,
    num_epochs: int = 3,
    learning_rate: float = 3e-4,
    cutoff_len: int = 512,
    val_set_size: float = 0.1,
    # lora hyperparams
    lora: bool = True,
    lora_r: int = 8,
    lora_alpha: int = 16,
    lora_dropout: float = 0.05,
    lora_target_modules: List[str] = ["query_key_value"],
    # llm hyperparams
    train_on_inputs: bool = False,  # if False, masks out inputs in loss
    add_eos_token: bool = False,
    group_by_length: bool = True,  # faster, but produces an odd training loss curve
    resume_from_checkpoint: str = None,  # either training checkpoint or final adapter
):
    if int(os.environ.get("LOCAL_RANK", 0)) == 0:
        print(
            f"Training LoRA model with params:\n"
            f"base_model: {base_model}\n"
            f"data_path: {data_path}\n"
            f"output_dir: {output_dir}\n"
            f"task_name: {task_name}\n"
            f"prompt_lang: {prompt_lang}\n"
            f"load_8bit: {load_8bit}\n"
            f"batch_size: {batch_size}\n"
            f"micro_batch_size: {micro_batch_size}\n"
            f"num_epochs: {num_epochs}\n"
            f"learning_rate: {learning_rate}\n"
            f"cutoff_len: {cutoff_len}\n"
            f"val_set_size: {val_set_size}\n"
            f"lora: {lora}\n"
            f"lora_r: {lora_r}\n"
            f"lora_alpha: {lora_alpha}\n"
            f"lora_dropout: {lora_dropout}\n"
            f"lora_target_modules: {lora_target_modules}\n"
            f"train_on_inputs: {train_on_inputs}\n"
            f"add_eos_token: {add_eos_token}\n"
            f"group_by_length: {group_by_length}\n"
            f"resume_from_checkpoint: {resume_from_checkpoint or False}\n"
        )
    assert (
        base_model
    ), "Please specify a --base_model, e.g. --base_model='huggyllama/llama-7b'"
    gradient_accumulation_steps = batch_size // micro_batch_size

    prompter = Prompter(task_name=task_name, prompt_lang=prompt_lang)

    device_map = "auto"
    world_size = int(os.environ.get("WORLD_SIZE", 1))
    ddp = world_size != 1
    if ddp:
        device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)}
        gradient_accumulation_steps = gradient_accumulation_steps // world_size

    # Check if parameter passed or if set within environ
    use_wandb = "WANDB_PROJECT" in os.environ and len(os.environ["WANDB_PROJECT"]) > 0

    model = AutoModelForCausalLM.from_pretrained(
        base_model,
        load_in_8bit=load_8bit,
        torch_dtype=torch.float16,
        device_map=device_map,
        resume_download=True
    )

    if load_8bit:
        model = prepare_model_for_int8_training(model)

    if lora:
        config = LoraConfig(
            r=lora_r,
            lora_alpha=lora_alpha,
            target_modules=lora_target_modules,
            lora_dropout=lora_dropout,
            bias="none",
            task_type="CAUSAL_LM",
        )
        model = get_peft_model(model, config)

    tokenizer = AutoTokenizer.from_pretrained(base_model)

    tokenizer.pad_token_id = (
        0  # unk. we want this to be different from the eos token
    )
    tokenizer.padding_side = "left"  # Allow batched inference

    def tokenize(prompt, add_eos_token=True):
        # there's probably a way to do this with the tokenizer settings
        # but again, gotta move fast
        result = tokenizer(
            prompt,
            truncation=True,
            max_length=cutoff_len,
            padding=False,
            return_tensors=None,
        )
        if (
            result["input_ids"][-1] != tokenizer.eos_token_id
            and len(result["input_ids"]) < cutoff_len
            and add_eos_token
        ):
            result["input_ids"].append(tokenizer.eos_token_id)
            result["attention_mask"].append(1)

        result["labels"] = result["input_ids"].copy()

        return result

    def generate_and_tokenize_prompt(data_point):
        full_prompt = prompter.get_prompt(data_point, with_label=True)
        tokenized_full_prompt = tokenize(full_prompt)
        if not train_on_inputs:
            user_prompt = prompter.get_prompt(data_point, with_label=False)
            tokenized_user_prompt = tokenize(
                user_prompt, add_eos_token=add_eos_token
            )
            user_prompt_len = len(tokenized_user_prompt["input_ids"])

            if add_eos_token:
                user_prompt_len -= 1

            tokenized_full_prompt["labels"] = [
                -100
            ] * user_prompt_len + tokenized_full_prompt["labels"][
                user_prompt_len:
            ]  # could be sped up, probably
        return tokenized_full_prompt

    if lora and resume_from_checkpoint:
        # Check the available weights and load them
        checkpoint_name = os.path.join(
            resume_from_checkpoint, "pytorch_model.bin"
        )  # Full checkpoint
        if not os.path.exists(checkpoint_name):
            checkpoint_name = os.path.join(
                resume_from_checkpoint, "adapter_model.bin"
            )  # only LoRA model - LoRA config above has to fit
            resume_from_checkpoint = (
                False  # So the trainer won't try loading its state
            )
        # The two files above have a different name depending on how they were saved, but are actually the same.
        if os.path.exists(checkpoint_name):
            print(f"Restarting from {checkpoint_name}")
            adapters_weights = torch.load(checkpoint_name)
            set_peft_model_state_dict(model, adapters_weights)
        else:
            print(f"Checkpoint {checkpoint_name} not found")

    if lora:
        model.print_trainable_parameters()  # Be more transparent about the % of trainable params.

    data = load_dataset("json", data_files=data_path)
    remove_columns = data["train"].column_names
    if val_set_size > 0:
        train_val = data["train"].train_test_split(
            test_size=val_set_size, shuffle=True, seed=42
        )
        train_data = (
            train_val["train"].shuffle().map(generate_and_tokenize_prompt, remove_columns=remove_columns)
        )
        val_data = (
            train_val["test"].shuffle().map(generate_and_tokenize_prompt, remove_columns=remove_columns)
        )
    else:
        train_data = data.shuffle().map(generate_and_tokenize_prompt, remove_columns=remove_columns)
        val_data = None

    if not ddp and torch.cuda.device_count() > 1:
        # keeps Trainer from trying its own DataParallelism when more than 1 gpu is available
        model.is_parallelizable = True
        model.model_parallel = True

    trainer = transformers.Trainer(
        model=model,
        train_dataset=train_data,
        eval_dataset=val_data,
        args=transformers.TrainingArguments(
            per_device_train_batch_size=micro_batch_size,
            gradient_accumulation_steps=gradient_accumulation_steps,
            warmup_ratio=0.1,
            num_train_epochs=num_epochs,
            learning_rate=learning_rate,
            fp16=True,
            logging_steps=10,
            optim="adamw_torch",
            evaluation_strategy="steps" if val_set_size > 0 else "no",
            save_strategy="steps",
            eval_steps=50 if val_set_size > 0 else None,
            save_steps=50,
            output_dir=output_dir,
            save_total_limit=1,
            load_best_model_at_end=True if val_set_size > 0 else False,
            ddp_find_unused_parameters=False if ddp else None,
            group_by_length=group_by_length,
            report_to="wandb" if use_wandb else None
        ),
        data_collator=transformers.DataCollatorForSeq2Seq(
            tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True
        ),
    )

    model.config.use_cache = False

    if lora:
        # such that only save lora weights in the intermediate checkpoints
        old_state_dict = model.state_dict
        model.state_dict = (
            lambda self, *_, **__: get_peft_model_state_dict(
                self, old_state_dict()
            )
        ).__get__(model, type(model))

    if torch.__version__ >= "2" and sys.platform != "win32":
        model = torch.compile(model)

    trainer.train(resume_from_checkpoint=resume_from_checkpoint)

    if lora:
        # peft will detect lora weights automatically, so here we convert back to full state_dict
        model.state_dict = old_state_dict
        model.save_pretrained(output_dir)
    else:
        model.save_pretrained(output_dir)
        tokenizer.save_pretrained(output_dir)
    
    print(
        "\n If there's a warning about missing keys above, please disregard :)"
    )


if __name__ == "__main__":
    fire.Fire(train)