#
# Copyright 2016 The BigDL Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Some parts of this file is adapted from
# https://github.com/tloen/alpaca-lora/blob/main/finetune.py
#
# Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li

# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import os
from typing import List

import torch
import transformers
from datasets import load_dataset
import accelerate
from typing import Union
from datetime import datetime

from transformers import AutoTokenizer
from peft import (
    get_peft_model_state_dict,
    set_peft_model_state_dict,
)

from ipex_llm.transformers import AutoModelForCausalLM
# import them from ipex_llm.transformers.qlora to get a IPEX-LLM compatible Peft model
from ipex_llm.transformers.qlora import get_peft_model, prepare_model_for_kbit_training,\
    LoraConfig
import json

local_rank = 0
world_size = 1

class Prompter(object):
    __slots__ = ("data")

    def __init__(self, data_path: str = ""):
        self.data = None
        if not os.path.exists(data_path):
            raise Exception(f"Can't read {data_path}")
        with open(data_path) as fp:
            self.data = json.load(fp)

    def generate_prompt(
        self,
        instruction: str,
        input: Union[None, str]=None,
        label: Union[None, str]=None,
    ) -> str:
        # returns the full prompt from instruction and optional input
        # if a label (=response, =output) is provided, it's also appended.
        if input:
            ret = f"### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n"
        else:
            ret = f"### Instruction:\n{instruction}\n\n### Response:\n"
        if label:
            ret = f"{ret}{label}"
        return ret

    def get_response(self, output: str) -> str:
        return output.split("### Response:")[1].strip()

"""Data processing to get train data and val data"""
def tokenize_qa(prompt, tokenizer, max_length=1, 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=max_length,
        padding=False,
        return_tensors=None,
    )
    if (
        result["input_ids"][-1] != tokenizer.eos_token_id
        and len(result["input_ids"]) < max_length
        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(entry, prompter, tokenizer, max_length, add_eos_token, train_on_inputs):
    full_prompt = prompter.generate_prompt(
        entry["question"],
        None,
        entry["answer"],
    )
    print({ 'full_prompt:' full_prompt })
    tokenized_full_prompt = tokenize_qa(full_prompt, tokenizer, max_length=max_length)
    if not train_on_inputs:
        user_prompt = prompter.generate_prompt(
            entry["question"], None
        )
        tokenized_user_prompt = tokenize_qa(
            user_prompt, tokenizer, add_eos_token=add_eos_token, max_length=max_length
        )
        user_prompt_len = len(tokenized_user_prompt["input_ids"])
        if add_eos_token:
            user_prompt_len -= 1
        # Mask out the input from the labels
        tokenized_full_prompt["labels"] = [
            -100
        ] * user_prompt_len + tokenized_full_prompt["labels"][
            user_prompt_len:
        ]  # could be sped up, probably
    return tokenized_full_prompt


def train(
    # model/data params
    base_model: str = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B",
    data_path: str = "./results/doc/resume/generic.txt.qa.json",
    output_dir: str = "./bigdl-qlora-alpaca",
    # training hyperparams
    bf16: bool = True,  # default to bf16
    batch_size: int = 128,
    micro_batch_size: int = 2,  # default to be 2, limited by GPU memory
    num_epochs: int = 3,
    learning_rate: float = 3e-5,  # default to be 3e-5 to avoid divergence
    cutoff_len: int = 256,
    val_set_size: int = 2000,
    # lora hyperparams
    lora_r: int = 8,
    lora_alpha: int = 16,
    lora_dropout: float = 0.05,
    lora_target_modules: List[str] = [
        "q_proj",
        "v_proj",
        "k_proj",
        "o_proj",
        "up_proj",
        "down_proj",
        "gate_proj"
    ],
    # llm hyperparams
    train_on_inputs: bool = True,  # 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
    resume_from_checkpoint: str = "",  # either training checkpoint or final adapter
    prompt_template_name: str = "alpaca",  # The prompt template to use, will default to alpaca.
    gradient_checkpointing: bool = False,
    training_mode: str = "lora",
    save_checkpoint: bool = True,
):
    if training_mode != "lora":
        raise Exception(f"This example is for lora training mode, but got training_mode={training_mode}.")
    
    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"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"
        f"prompt template: {prompt_template_name}\n"
        f"training_mode: {training_mode}\n"
        f"save_checkpoint: {save_checkpoint}\n"
    )
    
    gradient_accumulation_steps = batch_size // micro_batch_size

    prompter = Prompter("./results/doc/resume/generic.txt.qa.json")

    model = AutoModelForCausalLM.from_pretrained(
        base_model,
#	    load_in_4bit=True,
        load_in_low_bit="bf16",
        optimize_model=False,
#       optimize_model=True,
        torch_dtype=torch.bfloat16,
        modules_to_not_convert=["lm_head"],
        trust_remote_code=True,
    )

#    model = model.half()

    tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)

    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    # Prepare a IPEX-LLM compatible Peft model
    model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=gradient_checkpointing)

    config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        target_modules=lora_target_modules,
        lora_dropout=lora_dropout,
        bias="none",
        task_type="CAUSAL_LM",
        training_mode=training_mode,
    )
    print(f"Lora Config: {config}")
    model = get_peft_model(model, config)

    dataset = load_dataset("json", data_files=data_path)
    tokenized_dataset = dataset.map(generate_and_tokenize_prompt, fn_kwargs={
        "prompter": prompter,
        "tokenizer": tokenizer,
        "max_length": cutoff_len,
        "add_eos_token": add_eos_token,
        "train_on_inputs": train_on_inputs
    }, batched=True)
    model.print_trainable_parameters()  # Be more transparent about the % of trainable params.

    split = tokenized_dataset.train_test_split(test_size=0.2)
    train_dataset = split["train"]
    eval_dataset = split["test"]

    trainer = transformers.Trainer(
        model=model,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        args=transformers.TrainingArguments(
            per_device_train_batch_size=micro_batch_size,
            gradient_accumulation_steps=gradient_accumulation_steps,
            # warmup_ratio=0.03,
            # warmup_steps=100,
            max_grad_norm=0.3,
            num_train_epochs=num_epochs,
            learning_rate=learning_rate,
            lr_scheduler_type="cosine",
            bf16=True,  # ensure training more stable
            logging_steps=1,
            optim="adamw_torch",
            evaluation_strategy="steps" if val_set_size > 0 else "no",
            save_strategy="steps" if save_checkpoint else "no",
            eval_steps=100 if val_set_size > 0 else None,
            save_steps=100,
            output_dir=output_dir,
            save_total_limit=100,
            load_best_model_at_end=True if val_set_size > 0 and save_checkpoint else False,
            group_by_length=group_by_length,
            gradient_checkpointing=gradient_checkpointing,
            ddp_backend="ccl",
            save_safetensors=False,
        ),
        data_collator=transformers.DataCollatorForSeq2Seq(
            tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True
        ),
    )
    model.config.use_cache = False

    trainer.train(data_path="./results/doc/resume/generic.txt.qa.json", resume_from_checkpoint=resume_from_checkpoint)

    model.save_pretrained(output_dir)

    print(
        "\n If there's a warning about missing keys above, please disregard :)"
    )


if __name__ == "__main__":
    train()