finetuneSFTR1.py

import os
import json
from typing import List

import fire
import torch
import transformers
from datasets import load_dataset

from peft import (
    LoraConfig,
    get_peft_model,
    get_peft_model_state_dict,
    prepare_model_for_kbit_training,
    set_peft_model_state_dict,
)
from modelscope import AutoModelForCausalLM, AutoTokenizer

device = "cuda:0"  # the device to load the model onto

def train(
    #base_model: str = "Llama-3.2-1B-Instruct",
    #base_model: str = "Llama-3.2-3B-Instruct",
    #base_model: str = "Qwen2.5-0.5B-Instruct",
    #base_model: str = "Qwen2.5-1.5B-Instruct",
    base_model: str = "DeepSeek-R1-Distill-Qwen-1.5B",

    #data_path: str = "./data/ringo1-CoT_demo.json",
    #data_path: str = "./data/openr1-SFT.json",
    #data_path: str = "./data/magpie-r1.json",
    data_path: str = "./data/alpaca_r1_data_zh-remote.json",
    #data_path: str = "./data/alpaca_r1_data_zh-localpost.json",

    output_dir: str = "./lora-out",
    batch_size: int = 5,
    micro_batch_size: int = 4,
    num_epochs: int = 1,
    learning_rate: float = 3e-4,
    cutoff_len: int = 1024,
    val_set_size: int = 1,
    lora_r: int = 8,
    lora_alpha: int = 16,
    lora_dropout: float = 0.05,

    # lora_target_modules: List[str] = [
    #     "q_proj",
    #     "k_proj",
    # ],

    # lora_target_modules: List[str] = [
    #     "q_proj",
    #     "k_proj",
    #     "v_proj",
    #     "o_proj",
    # ],

    lora_target_modules: List[str] = ['q_proj', 'k_proj', 'v_proj', 'o_proj', "gate_proj", "up_proj", "down_proj"],
    
    train_on_inputs: bool = True,
    group_by_length: bool = False,
    wandb_project: str = "",
    wandb_run_name: str = "",
    wandb_watch: str = "",
    wandb_log_model: str = "",
    resume_from_checkpoint: str = None,
):
    print(
        f"Training Alpaca-LoRA model with params:\n"
        f"base_model: {base_model}\n"
        f"data_path: {data_path}\n"
        f"output_dir: {output_dir}\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_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"group_by_length: {group_by_length}\n"
        f"wandb_project: {wandb_project}\n"
        f"wandb_run_name: {wandb_run_name}\n"
        f"wandb_watch: {wandb_watch}\n"
        f"wandb_log_model: {wandb_log_model}\n"
        f"resume_from_checkpoint: {resume_from_checkpoint}\n"
    )

    gradient_accumulation_steps = batch_size // micro_batch_size

    use_wandb = len(wandb_project) > 0 or (
        "WANDB_PROJECT" in os.environ and len(os.environ["WANDB_PROJECT"]) > 0
    )
    if len(wandb_project) > 0:
        os.environ["WANDB_PROJECT"] = wandb_project
    if len(wandb_watch) > 0:
        os.environ["WANDB_WATCH"] = wandb_watch
    if len(wandb_log_model) > 0:
        os.environ["WANDB_LOG_MODEL"] = wandb_log_model

    model = AutoModelForCausalLM.from_pretrained(
        base_model,
        load_in_8bit=False,
        torch_dtype=torch.float16,
        device_map=device,
    )

    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):
        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 = generate_prompt(data_point)
        tokenized_full_prompt = tokenize(full_prompt)
        if not train_on_inputs:
            user_prompt = generate_prompt({**data_point, "output": ""})
            tokenized_user_prompt = tokenize(user_prompt, add_eos_token=False)
            user_prompt_len = len(tokenized_user_prompt["input_ids"])
            tokenized_full_prompt["labels"] = [
                -100
            ] * user_prompt_len + tokenized_full_prompt["labels"][user_prompt_len:]
        return tokenized_full_prompt

    model = prepare_model_for_kbit_training(model)

    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)

    if data_path.endswith(".json"):
        data = load_dataset("json", data_files=data_path)
    else:
        data = load_dataset(data_path)

    if resume_from_checkpoint:
        checkpoint_name = os.path.join(
            resume_from_checkpoint, "adapter_model.bin"
        )
        if os.path.exists(checkpoint_name):
            print(f"Restarting from {checkpoint_name}")
            adapters_weights = torch.load(checkpoint_name)
            model = set_peft_model_state_dict(model, adapters_weights)
        else:
            print(f"Checkpoint {checkpoint_name} not found")

    model.print_trainable_parameters()

    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)
        val_data = train_val["test"].shuffle().map(generate_and_tokenize_prompt)
    else:
        train_data = data["train"].shuffle().map(generate_and_tokenize_prompt)
        val_data = None

    from trl import SFTTrainer,SFTConfig
    trainer = SFTTrainer(
        model=model,
        train_dataset=train_data,
        eval_dataset=val_data,
        args=transformers.TrainingArguments( #SFTConfig
            per_device_train_batch_size=micro_batch_size,
            gradient_accumulation_steps=gradient_accumulation_steps,
            warmup_steps=100,
            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=200 if val_set_size > 0 else None,
            save_steps=200,
            output_dir=output_dir,
            save_total_limit=3,
            load_best_model_at_end=True if val_set_size > 0 else False,
            ddp_find_unused_parameters=False,
            group_by_length=group_by_length,
            report_to="wandb" if use_wandb else None,
            run_name=wandb_run_name 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

    trainer.train(resume_from_checkpoint=resume_from_checkpoint)

    # Save the LoRA adapter
    lora_weights = get_peft_model_state_dict(model)
    torch.save(lora_weights, os.path.join(output_dir, "adapter_model.bin"))

    config_dict = config.to_dict()
    config_dict["target_modules"] = list(config_dict["target_modules"])
    with open(os.path.join(output_dir, "adapter_config.json"), "w") as f:
        json.dump(config_dict, f, indent=4)

    print("\nTraining complete. LoRA adapter weights and config saved.")


def generate_prompt(data_point):
    return f"""Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{data_point["instruction"]}

### Response:
{data_point["output"]}"""

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