finetune.py

#!/bin/python3

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
#

import os
import sys
from typing import List
import datetime

import fire
import torch
import transformers
from datasets import load_dataset
import wandb

from transformers import (
    TrainerCallback,
    TrainingArguments,
    TrainerState,
    TrainerControl,
)
from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
from transformers import LlamaForCausalLM, LlamaTokenizer

from peft import (
    LoraConfig,
    get_peft_model,
    prepare_model_for_int8_training,
    set_peft_model_state_dict,
)

from chatbots.conversation import Conversation
from src.utils import fc_prefix, fc_suffix


class SavePeftModelCallback(TrainerCallback):
    def on_save(
        self,
        args: TrainingArguments,
        state: TrainerState,
        control: TrainerControl,
        **kwargs,
    ):
        checkpoint_folder = os.path.join(
            args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}"
        )

        kwargs["model"].save_pretrained(checkpoint_folder)

        pytorch_model_path = os.path.join(checkpoint_folder, "pytorch_model.bin")
        torch.save({}, pytorch_model_path)
        return control


class LoadBestPeftModelCallback(TrainerCallback):
    def on_train_end(
        self,
        args: TrainingArguments,
        state: TrainerState,
        control: TrainerControl,
        **kwargs,
    ):
        print(
            f"Loading best peft model from {state.best_model_checkpoint} (score: {state.best_metric})."
        )
        best_model_path = os.path.join(state.best_model_checkpoint, "adapter_model.bin")
        adapters_weights = torch.load(best_model_path)
        model = kwargs["model"]
        set_peft_model_state_dict(model, adapters_weights)
        return control


def train(
    # model/data params
    base_model: str = "",
    data_path: str = "",
    output_dir: str = "",
    # training hyperparams
    batch_size: int = 128,
    micro_batch_size: int = 8,
    num_epochs: int = 1,
    learning_rate: float = 3e-4,
    cutoff_len: int = 4096,
    val_set_size: int = 0,
    eval_steps: int = 200,
    save_steps: int = 1000,
    lr_scheduler: str = "cosine",
    warmup_steps: int = 100,
    # lora hyperparams
    lora_r: int = 16,
    lora_alpha: int = 16,
    lora_dropout: float = 0.05,
    # from peft docs: ["q_proj", "k_proj", "v_proj", "o_proj", "fc_in", "fc_out", "wte", "gate_proj", "down_proj", "up_proj"]
    lora_target_modules: List[str] = ["gate_proj", "down_proj", "up_proj"],
    # llm hyperparams
    train_on_response: bool = False,  # if False, masks out inputs in loss
    add_eos_token: bool = False,
    group_by_length: bool = False,  # faster, but produces an odd training loss curve
    # wandb params
    wandb_project: str = "",
    wandb_run_name: str = "",
    wandb_watch: str = "",  # options: false | gradients | all
    wandb_log_model: str = "",  # options: false | true
    resume_from_checkpoint: str = None,  # either training checkpoint or final adapter
):
    if int(os.environ.get("LOCAL_RANK", 0)) == 0:
        print(
            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"eval_steps: {eval_steps}\n"
            f"save_steps: {save_steps}\n"
            f"lr_scheduler: {lr_scheduler}\n"
            f"warmup_steps: {warmup_steps}\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_response: {train_on_response}\n"
            f"add_eos_token: {add_eos_token}\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 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

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

    # Check if parameter passed or if set within environ
    use_wandb = len(wandb_project) > 0 or (
        "WANDB_PROJECT" in os.environ and len(os.environ["WANDB_PROJECT"]) > 0
    )
    # Only overwrite environ if wandb param passed
    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 = LlamaForCausalLM.from_pretrained(
        base_model, load_in_8bit=True, torch_dtype=torch.float16, device_map=device_map
    )

    tokenizer = LlamaTokenizer.from_pretrained(base_model)

    bos = tokenizer.bos_token_id
    eos = tokenizer.eos_token_id
    pad = tokenizer.pad_token_id
    print(
        "pre-trained model's BOS EOS and PAD token id:",
        bos,
        eos,
        pad,
        " => It should be 1 2 None",
    )

    tokenizer.pad_token_id = 0  # unk. we want this to be different from the eos token
    tokenizer.padding_side = "right"

    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_conversation(data_point):
        # data in the conversation fashion
        system_prompt = data_point["system"]
        conversation = data_point["conversation"]

        labels = []
        tokenized_ids = []
        target_len = 0

        # system message part
        tokenized_system_prompt = tokenize(system_prompt, add_eos_token=add_eos_token)
        system_prompt_len = len(tokenized_system_prompt["input_ids"])
        if add_eos_token:
            system_prompt_len -= 1
        tokenized_ids += tokenized_system_prompt["input_ids"]
        labels += [-100] * system_prompt_len

        # conversation
        for message in conversation:
            ###### user turn ######
            if message["role"] == "user":
                user_input = message["content"]
                tokenized_user_input = tokenize(user_input, add_eos_token=add_eos_token)
                user_input_len = len(tokenized_user_input["input_ids"])
                if add_eos_token:
                    user_input_len -= 1
                tokenized_ids += tokenized_user_input["input_ids"]
                labels += [-100] * user_input_len

            ###### assistant turn ######
            elif message["role"] == "assistant":
                assistant_output = message["content"]

                ### select target output ###
                if fc_prefix in assistant_output and fc_suffix in assistant_output:
                    function_call = assistant_output.split(fc_suffix)[0] + fc_suffix
                    response = assistant_output.split(fc_suffix)[1]
                else:
                    function_call = ""
                    response = assistant_output

                ### part 1: function call ###
                if function_call:
                    tokenized_function_call = tokenize(
                        function_call, add_eos_token=add_eos_token
                    )
                    function_call_output_len = len(tokenized_function_call["input_ids"])
                    if add_eos_token:
                        function_call_output_len -= 1
                    tokenized_ids += tokenized_function_call["input_ids"]
                    labels += tokenized_function_call["input_ids"]
                    target_len += function_call_output_len

                ### part 2: response ###
                if response:
                    tokenized_response = tokenize(response, add_eos_token=add_eos_token)
                    response_output_len = len(tokenized_response["input_ids"])
                    if add_eos_token:
                        response_output_len -= 1
                    tokenized_ids += tokenized_response["input_ids"]
                    if train_on_response:
                        labels += tokenized_response["input_ids"]
                        target_len += response_output_len
                    else:
                        labels += [-100] * response_output_len

        assert len(tokenized_ids) == len(labels)
        return {
            "input_ids": torch.LongTensor(tokenized_ids),
            "labels": torch.LongTensor(labels),
        }

    model = prepare_model_for_int8_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") or data_path.endswith(".jsonl"):
        data = load_dataset("json", data_files=data_path)
    else:
        data = load_dataset(data_path)

    if 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")

    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_conversation)
        )
        val_data = train_val["test"].shuffle().map(generate_and_tokenize_conversation)
    else:
        train_data = data["train"].shuffle().map(generate_and_tokenize_conversation)
        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,
            per_device_eval_batch_size=micro_batch_size,
            gradient_accumulation_steps=gradient_accumulation_steps,
            warmup_steps=warmup_steps,
            num_train_epochs=num_epochs,
            learning_rate=learning_rate,
            # dataloader_num_workers=16,
            fp16=True,
            logging_steps=1,
            optim="adamw_torch",
            evaluation_strategy="steps" if val_set_size > 0 else "no",
            save_strategy="steps",
            eval_steps=eval_steps if val_set_size > 0 else None,
            save_steps=save_steps,
            lr_scheduler_type=lr_scheduler,
            output_dir=output_dir,
            save_total_limit=2,
            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,
            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
        ),
        # callbacks=[SavePeftModelCallback, LoadBestPeftModelCallback], # ONLY USE LoadBestPeftModelCallback if val_set_size > 0
    )
    model.config.use_cache = False

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

    trainer.train(resume_from_checkpoint=resume_from_checkpoint)

    model.save_pretrained(output_dir)
    # model.base_model.save_pretrained(output_dir)
    pytorch_model_path = os.path.join(output_dir, "pytorch_model.bin")
    torch.save({}, pytorch_model_path)
    tokenizer.save_pretrained(output_dir)


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