finetune.py

import copy
import json
import os
import pathlib
from dataclasses import dataclass, field
from typing import Optional, Dict, Sequence

# import deepspeed
# deepspeed.ops.op_builder.CPUAdamBuilder().load()

import tokenizers
import torch
import transformers
from PIL import Image
from torch.utils.data import Dataset
from transformers import LlamaConfig

from llava.constants import DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, IGNORE_INDEX
from llava.mm_utils import tokenizer_image_token
from llava.model.language_model.llava_llama import LlavaLlamaForCausalLM
from llavanext.model.language_model.llava_llama import LlavaLlamaForCausalLM as LlavaLlamaForCausalLMNext
from llava import conversation as conversation_lib
from llava.train.llava_trainer import LLaVATrainer, maybe_zero_3

local_rank = 0




def rank0_print(*args):
    if local_rank == 0:
        print(*args)


from packaging import version
IS_TOKENIZER_GREATER_THAN_0_14 = version.parse(tokenizers.__version__) >= version.parse('0.14')


@dataclass
class ModelArguments:
    model_name_or_path: Optional[str] = field(default='liuhaotian/llava-v1.5-7b')
    version: Optional[str] = field(default="v1")
    freeze_backbone: bool = field(default=False)
    tune_mm_mlp_adapter: bool = field(default=False)
    vision_tower: Optional[str] = field(default='openai/clip-vit-large-patch14-336')
    mm_vision_select_layer: Optional[int] = field(default=-2)  # default to the last layer
    pretrain_mm_mlp_adapter: Optional[str] = field(default=None)
    mm_projector_type: Optional[str] = field(default='linear')
    mm_use_im_start_end: bool = field(default=False)
    mm_use_im_patch_token: bool = field(default=False)
    mm_patch_merge_type: Optional[str] = field(default='flat')
    mm_vision_select_feature: Optional[str] = field(default="patch")


@dataclass
class DataArguments:
    data_path: str = field(default='/home/glr/code/LLaMA-Factory/data/mllm_demo.json',
                           metadata={"help": "Path to the training data."})
    lazy_preprocess: bool = False
    is_multimodal: bool = False
    image_folder: Optional[str] = field(default=None)
    image_aspect_ratio: str = 'square'


@dataclass
class TrainingArguments(transformers.TrainingArguments):
    cache_dir: Optional[str] = field(default=None)
    optim: str = field(default="adamw_torch")
    remove_unused_columns: bool = field(default=False)
    freeze_mm_mlp_adapter: bool = field(default=False)
    # mpt_attn_impl: Optional[str] = field(default="triton")
    model_max_length: int = field(
        default=512,
        metadata={
            "help":
                "Maximum sequence length. Sequences will be right padded (and possibly truncated)."
        },
    )
    # double_quant: bool = field(
    #     default=True,
    #     metadata={"help": "Compress the quantization statistics through double quantization."}
    # )
    # quant_type: str = field(
    #     default="nf4",
    #     metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."}
    # )
    bits: int = field(
        default=16,
        metadata={"help": "How many bits to use."}
    )
    lora_enable: bool = True
    lora_r: int = 64
    lora_alpha: int = 16
    lora_dropout: float = 0.05
    lora_weight_path: str = ""
    lora_bias: str = "none"
    lora_layer: str = "k_proj, gate_proj, q_proj, up_proj, o_proj, v_proj, down_proj"
    mm_projector_lr: Optional[float] = None
    group_by_modality_length: bool = field(default=False)
    per_device_train_batch_size: int = 1
    gradient_accumulation_steps: int = 16
    deepspeed: str = 'zero3.json'


def find_all_linear_names(model):
    cls = torch.nn.Linear
    lora_module_names = set()
    multimodal_keywords = ['mm_projector', 'vision_tower', 'vision_resampler']
    for name, module in model.named_modules():
        if any(mm_keyword in name for mm_keyword in multimodal_keywords):
            continue
        if isinstance(module, cls):
            names = name.split('.')
            lora_module_names.add(names[0] if len(names) == 1 else names[-1])
    # print(lora_module_names)
    if 'lm_head' in lora_module_names:  # needed for 16-bit
        lora_module_names.remove('lm_head')
    print('#'*1000)
    print('lora_module_names', lora_module_names)
    return list(lora_module_names)


def preprocess_multimodal(
    sources: Sequence[str],
    data_args: DataArguments
) -> Dict:
    is_multimodal = data_args.is_multimodal
    if not is_multimodal:
        return sources

    for source in sources:
        for sentence in source:
            if DEFAULT_IMAGE_TOKEN in sentence['value']:
                sentence['value'] = sentence['value'].replace(DEFAULT_IMAGE_TOKEN, '').strip()
                sentence['value'] = DEFAULT_IMAGE_TOKEN + '\n' + sentence['value']
                sentence['value'] = sentence['value'].strip()
                if "mmtag" in conversation_lib.default_conversation.version:
                    sentence['value'] = sentence['value'].replace(DEFAULT_IMAGE_TOKEN, '<Image>' + DEFAULT_IMAGE_TOKEN + '</Image>')
            replace_token = DEFAULT_IMAGE_TOKEN
            if data_args.mm_use_im_start_end:
                replace_token = DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN
            sentence["value"] = sentence["value"].replace(DEFAULT_IMAGE_TOKEN, replace_token)

    return sources


def preprocess_v1(
    sources,
    tokenizer: transformers.PreTrainedTokenizer,
    has_image: bool = False
) -> Dict:
    conv = conversation_lib.default_conversation.copy()
    roles = {"human": conv.roles[0], "gpt": conv.roles[1]}

    # Apply prompt templates
    conversations = []
    for i, source in enumerate(sources):
        if roles[source[0]["from"]] != conv.roles[0]:
            # Skip the first one if it is not from human
            source = source[1:]

        conv.messages = []
        for j, sentence in enumerate(source):
            role = roles[sentence["from"]]
            assert role == conv.roles[j % 2], f"{i}"
            conv.append_message(role, sentence["value"])
        conversations.append(conv.get_prompt())

    # Tokenize conversations

    if has_image:
        input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
    else:
        input_ids = tokenizer(
            conversations,
            return_tensors="pt",
            padding="longest",
            max_length=tokenizer.model_max_length,
            truncation=True,
        ).input_ids

    targets = input_ids.clone()

    assert conv.sep_style == conversation_lib.SeparatorStyle.TWO

    # Mask targets
    sep = conv.sep + conv.roles[1] + ": "
    for conversation, target in zip(conversations, targets):
        total_len = int(target.ne(tokenizer.pad_token_id).sum())

        rounds = conversation.split(conv.sep2)
        cur_len = 1
        target[:cur_len] = IGNORE_INDEX
        for i, rou in enumerate(rounds):
            if rou == "":
                break

            parts = rou.split(sep)
            if len(parts) != 2:
                break
            parts[0] += sep

            if has_image:
                round_len = len(tokenizer_image_token(rou, tokenizer))
                instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 2
            else:
                round_len = len(tokenizer(rou).input_ids)
                instruction_len = len(tokenizer(parts[0]).input_ids) - 2

            if i != 0 and not tokenizer.legacy and IS_TOKENIZER_GREATER_THAN_0_14:
                round_len -= 1
                instruction_len -= 1

            target[cur_len : cur_len + instruction_len] = IGNORE_INDEX

            cur_len += round_len
        target[cur_len:] = IGNORE_INDEX

        if cur_len < tokenizer.model_max_length:
            if cur_len != total_len:
                target[:] = IGNORE_INDEX
                print(
                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
                    f" (ignored)"
                )

    return dict(
        input_ids=input_ids,
        labels=targets,
    )


def _add_speaker_and_signal(header, source, get_conversation=True):
    """Add speaker and start/end signal on each round."""
    BEGIN_SIGNAL = "### "
    END_SIGNAL = "\n"
    conversation = header
    for sentence in source:
        from_str = sentence["from"]
        if from_str.lower() == "human":
            from_str = conversation_lib.default_conversation.roles[0]
        elif from_str.lower() == "gpt":
            from_str = conversation_lib.default_conversation.roles[1]
        else:
            from_str = 'unknown'
        sentence["value"] = (BEGIN_SIGNAL + from_str + ": " +
                             sentence["value"] + END_SIGNAL)
        if get_conversation:
            conversation += sentence["value"]
    conversation += BEGIN_SIGNAL
    return conversation


def _tokenize_fn(strings: Sequence[str],
                 tokenizer: transformers.PreTrainedTokenizer) -> Dict:
    """Tokenize a list of strings."""
    tokenized_list = [
        tokenizer(
            text,
            return_tensors="pt",
            padding="longest",
            max_length=tokenizer.model_max_length,
            truncation=True,
        ) for text in strings
    ]
    input_ids = labels = [
        tokenized.input_ids[0] for tokenized in tokenized_list
    ]
    input_ids_lens = labels_lens = [
        tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item()
        for tokenized in tokenized_list
    ]
    return dict(
        input_ids=input_ids,
        labels=labels,
        input_ids_lens=input_ids_lens,
        labels_lens=labels_lens,
    )


def _mask_targets(target, tokenized_lens, speakers):
    # cur_idx = 0
    cur_idx = tokenized_lens[0]
    tokenized_lens = tokenized_lens[1:]
    target[:cur_idx] = IGNORE_INDEX
    for tokenized_len, speaker in zip(tokenized_lens, speakers):
        if speaker == "human":
            target[cur_idx+2:cur_idx + tokenized_len] = IGNORE_INDEX
        cur_idx += tokenized_len


def preprocess(
    sources: Sequence[str],
    tokenizer: transformers.PreTrainedTokenizer,
    has_image: bool = False
) -> Dict:
    """
    Given a list of sources, each is a conversation list. This transform:
    1. Add signal '### ' at the beginning each sentence, with end signal '\n';
    2. Concatenate conversations together;
    3. Tokenize the concatenated conversation;
    4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX.
    """
    # # 预训练时候调用的是plain，finetune时候调用的是v1
    # if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.PLAIN:
    #     # print('111')
    #     return preprocess_plain(sources, tokenizer)
    # if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.LLAMA_2:
    #     # print('222')
    #     return preprocess_llama_2(sources, tokenizer, has_image=has_image)
    if conversation_lib.default_conversation.version.startswith("v1"):
        # print('333')
        return preprocess_v1(sources, tokenizer, has_image=has_image)
    # if conversation_lib.default_conversation.version == "mpt":
    #     # print('444')
    #     return preprocess_mpt(sources, tokenizer, has_image=has_image)
    # # print('555')
    # add end signal and concatenate together
    conversations = []
    for source in sources:
        header = f"{conversation_lib.default_conversation.system}\n\n"
        conversation = _add_speaker_and_signal(header, source)
        conversations.append(conversation)
    # tokenize conversations
    def get_tokenize_len(prompts):
        return [len(tokenizer_image_token(prompt, tokenizer)) for prompt in prompts]

    if has_image:
        input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations]
    else:
        conversations_tokenized = _tokenize_fn(conversations, tokenizer)
        input_ids = conversations_tokenized["input_ids"]

    targets = copy.deepcopy(input_ids)
    for target, source in zip(targets, sources):
        if has_image:
            tokenized_lens = get_tokenize_len([header] + [s["value"] for s in source])
        else:
            tokenized_lens = _tokenize_fn([header] + [s["value"] for s in source], tokenizer)["input_ids_lens"]
        speakers = [sentence["from"] for sentence in source]
        _mask_targets(target, tokenized_lens, speakers)

    return dict(input_ids=input_ids, labels=targets)


class LazySupervisedDataset(Dataset):
    """Dataset for supervised fine-tuning."""

    def __init__(self, data_path: str,
                 tokenizer: transformers.PreTrainedTokenizer,
                 data_args: DataArguments):
        super(LazySupervisedDataset, self).__init__()
        list_data_dict = json.load(open(data_path, "r"))

        rank0_print("Formatting inputs...Skip in lazy mode")
        self.tokenizer = tokenizer
        self.list_data_dict = list_data_dict
        self.data_args = data_args

    def __len__(self):
        return len(self.list_data_dict)

    @property
    def lengths(self):
        length_list = []
        for sample in self.list_data_dict:
            img_tokens = 128 if 'image' in sample else 0
            length_list.append(sum(len(conv['value'].split()) for conv in sample['conversations']) + img_tokens)
        return length_list

    @property
    def modality_lengths(self):
        length_list = []
        for sample in self.list_data_dict:
            cur_len = sum(len(conv['value'].split()) for conv in sample['conversations'])
            cur_len = cur_len if 'image' in sample else -cur_len
            length_list.append(cur_len)
        return length_list

    def __getitem__(self, i) -> Dict[str, torch.Tensor]:
        sources = self.list_data_dict[i]
        if isinstance(i, int):
            sources = [sources]
        assert len(sources) == 1, "Don't know why it is wrapped to a list"  # FIXME
        if 'image' in sources[0]:
            image_file = self.list_data_dict[i]['image']
            image_folder = self.data_args.image_folder
            processor = self.data_args.image_processor
            image = Image.open(os.path.join(image_folder, image_file)).convert('RGB')
            if self.data_args.image_aspect_ratio == 'pad':
                def expand2square(pil_img, background_color):
                    width, height = pil_img.size
                    if width == height:
                        return pil_img
                    elif width > height:
                        result = Image.new(pil_img.mode, (width, width), background_color)
                        result.paste(pil_img, (0, (width - height) // 2))
                        return result
                    else:
                        result = Image.new(pil_img.mode, (height, height), background_color)
                        result.paste(pil_img, ((height - width) // 2, 0))
                        return result
                image = expand2square(image, tuple(int(x*255) for x in processor.image_mean))
                image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
            else:
                image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
            sources = preprocess_multimodal(
                copy.deepcopy([e["conversations"] for e in sources]),
                self.data_args)
        else:
            sources = copy.deepcopy([e["conversations"] for e in sources])
        data_dict = preprocess(
            sources,
            self.tokenizer,
            has_image=('image' in self.list_data_dict[i]))
        if isinstance(i, int):
            data_dict = dict(input_ids=data_dict["input_ids"][0],
                             labels=data_dict["labels"][0])

        # image exist in the data
        if 'image' in self.list_data_dict[i]:
            data_dict['image'] = image
        elif self.data_args.is_multimodal:
            # image does not exist in the data, but the model is multimodal
            crop_size = self.data_args.image_processor.crop_size
            data_dict['image'] = torch.zeros(3, crop_size['height'], crop_size['width'])
        return data_dict


class MergeSupervisedDataset(LazySupervisedDataset):
    def __init__(self, data_path: str,
                 tokenizer: transformers.PreTrainedTokenizer,
                 data_args: DataArguments):
        super(Dataset).__init__()
        data_path = data_path.split(',')
        self.list_data_dict = []
        for i in data_path:
            self.list_data_dict += json.load(open(i, "r"))

        rank0_print("Formatting inputs...Skip in lazy mode")
        self.tokenizer = tokenizer
        # self.list_data_dict = list_data_dict
        self.data_args = data_args


@dataclass
class DataCollatorForSupervisedDataset(object):
    """Collate examples for supervised fine-tuning."""

    tokenizer: transformers.PreTrainedTokenizer

    def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
        input_ids, labels = tuple([instance[key] for instance in instances]
                                  for key in ("input_ids", "labels"))
        input_ids = torch.nn.utils.rnn.pad_sequence(
            input_ids,
            batch_first=True,
            padding_value=self.tokenizer.pad_token_id)
        labels = torch.nn.utils.rnn.pad_sequence(labels,
                                                 batch_first=True,
                                                 padding_value=IGNORE_INDEX)
        input_ids = input_ids[:, :self.tokenizer.model_max_length]
        labels = labels[:, :self.tokenizer.model_max_length]
        batch = dict(
            input_ids=input_ids,
            labels=labels,
            attention_mask=input_ids.ne(self.tokenizer.pad_token_id),
        )

        if 'image' in instances[0]:
            images = [instance['image'] for instance in instances]
            if all(x is not None and x.shape == images[0].shape for x in images):
                batch['images'] = torch.stack(images)
            else:
                batch['images'] = images

        return batch



def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer,
                                data_args) -> Dict:
    """Make dataset and collator for supervised fine-tuning."""
    train_dataset = MergeSupervisedDataset(tokenizer=tokenizer,
                                           data_path=data_args.data_path,
                                           data_args=data_args)
    data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer)
    return dict(train_dataset=train_dataset,
                eval_dataset=None,
                data_collator=data_collator)


def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True):
    to_return = {k: t for k, t in named_params if "lora_" not in k}
    if require_grad_only:
        to_return = {k: t for k, t in to_return.items() if t.requires_grad}
    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
    return to_return


def get_peft_state_maybe_zero_3(named_params, bias):
    if bias == "none":
        to_return = {k: t for k, t in named_params if "lora_" in k}
    elif bias == "all":
        to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k}
    elif bias == "lora_only":
        to_return = {}
        maybe_lora_bias = {}
        lora_bias_names = set()
        for k, t in named_params:
            if "lora_" in k:
                to_return[k] = t
                bias_name = k.split("lora_")[0] + "bias"
                lora_bias_names.add(bias_name)
            elif "bias" in k:
                maybe_lora_bias[k] = t
        for k, t in maybe_lora_bias:
            if bias_name in lora_bias_names:
                to_return[bias_name] = t
    else:
        raise NotImplementedError
    to_return = {k: maybe_zero_3(v, ignore_status=True) for k, v in to_return.items()}
    return to_return


def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
    to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)}
    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
    return to_return

def safe_save_model_for_hf_trainer(trainer: transformers.Trainer,
                                   output_dir: str):
    """Collects the state dict and dump to disk."""

    if getattr(trainer.args, "tune_mm_mlp_adapter", False):
        # Only save Adapter
        keys_to_match = ['mm_projector']
        if getattr(trainer.args, "use_im_start_end", False):
            keys_to_match.extend(['embed_tokens', 'embed_in'])

        weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match)
        trainer.model.config.save_pretrained(output_dir)

        current_folder = output_dir.split('/')[-1]
        parent_folder = os.path.dirname(output_dir)
        if trainer.args.local_rank == 0 or trainer.args.local_rank == -1:
            if current_folder.startswith('checkpoint-'):
                mm_projector_folder = os.path.join(parent_folder, "mm_projector")
                os.makedirs(mm_projector_folder, exist_ok=True)
                torch.save(weight_to_save, os.path.join(mm_projector_folder, f'{current_folder}.bin'))
            else:
                torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin'))
        return

    if trainer.deepspeed:
        torch.cuda.synchronize()
        trainer.save_model(output_dir)
        return

    state_dict = trainer.model.state_dict()
    if trainer.args.should_save:
        cpu_state_dict = {
            key: value.cpu()
            for key, value in state_dict.items()
        }
        del state_dict
        trainer._save(output_dir, state_dict=cpu_state_dict)  # noqa


def train(attn_implementation=None):
    parser = transformers.HfArgumentParser(
        (ModelArguments, DataArguments, TrainingArguments))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    # 量化config
    bnb_model_from_pretrained_args = {}


    if 'llava-next' in model_args.model_name_or_path:
        class LlavaConfig(LlamaConfig):
            model_type = "llava_llama"
            temperature: float = 0.0  # reset to 0.0, previously 0.9 for Vicuna
            max_new_tokens: int = 1024
            do_sample: bool = False
            top_p: Optional[float] = None
            rope_scaling: Optional[dict] = {}

        lora_cfg_pretrained = LlavaConfig.from_pretrained(model_args.model_name_or_path,)
        model = LlavaLlamaForCausalLMNext.from_pretrained(
            model_args.model_name_or_path,
            low_cpu_mem_usage=True,
            config=lora_cfg_pretrained,
            attn_implementation=attn_implementation,
            # **kwargs
        )
    else:

        model = LlavaLlamaForCausalLM.from_pretrained(
            model_args.model_name_or_path,
            cache_dir=training_args.cache_dir,
            attn_implementation=attn_implementation,
            torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
            **bnb_model_from_pretrained_args
        )

    model.config.use_cache = False

    # 冻结backbone
    if model_args.freeze_backbone:
        model.model.requires_grad_(False)

    if training_args.gradient_checkpointing:
        model.enable_input_require_grads()

    if training_args.lora_enable:
        from peft import LoraConfig, get_peft_model
        lora_layer = training_args.lora_layer.replace(' ', '').split(',')
        rank0_print('Set lora layer', lora_layer)
        lora_config = LoraConfig(
            # 秩数
            r=training_args.lora_r,
            # self.scaling = self.lora_alpha / self.r
            # 对lora结果做个缩放
            lora_alpha=training_args.lora_alpha,
            # target_modules=find_all_linear_names(model),
            target_modules=lora_layer,
            lora_dropout=training_args.lora_dropout,
            bias=training_args.lora_bias,
            task_type="CAUSAL_LM",
        )
        if training_args.bits == 16:
            if training_args.bf16:
                model.to(torch.bfloat16)
            if training_args.fp16:
                model.to(torch.float16)
        rank0_print("Adding LoRA adapters...")
        model = get_peft_model(model, lora_config)

    # 加载tokenizer
    tokenizer = transformers.AutoTokenizer.from_pretrained(
        model_args.model_name_or_path,
        cache_dir=training_args.cache_dir,
        model_max_length=training_args.model_max_length,
        padding_side="right",
        use_fast=False,
    )
    tokenizer.pad_token = tokenizer.unk_token
    if model_args.version in conversation_lib.conv_templates:
        conversation_lib.default_conversation = conversation_lib.conv_templates[model_args.version]
    else:
        conversation_lib.default_conversation = conversation_lib.conv_templates["vicuna_v1"]

    model.get_model().initialize_vision_modules(
        model_args=model_args,
        fsdp=training_args.fsdp
    )

    vision_tower = model.get_vision_tower()
    vision_tower.to(dtype=torch.bfloat16 if training_args.bf16 else torch.float16, device=training_args.device)

    data_args.image_processor = vision_tower.image_processor
    data_args.is_multimodal = True

    model.config.image_aspect_ratio = data_args.image_aspect_ratio
    model.config.tokenizer_padding_side = tokenizer.padding_side
    model.config.tokenizer_model_max_length = tokenizer.model_max_length

    model.config.tune_mm_mlp_adapter = training_args.tune_mm_mlp_adapter = model_args.tune_mm_mlp_adapter
    if model_args.tune_mm_mlp_adapter:
        model.requires_grad_(False)
        for p in model.get_model().mm_projector.parameters():
            p.requires_grad = True

    model.config.freeze_mm_mlp_adapter = training_args.freeze_mm_mlp_adapter
    if training_args.freeze_mm_mlp_adapter:
        for p in model.get_model().mm_projector.parameters():
            p.requires_grad = False

    model.config.mm_use_im_start_end = data_args.mm_use_im_start_end = model_args.mm_use_im_start_end
    model.config.mm_projector_lr = training_args.mm_projector_lr
    training_args.use_im_start_end = model_args.mm_use_im_start_end
    model.config.mm_use_im_patch_token = model_args.mm_use_im_patch_token
    model.initialize_vision_tokenizer(model_args, tokenizer=tokenizer)

    data_module = make_supervised_data_module(tokenizer=tokenizer,
                                              data_args=data_args)

    trainer = LLaVATrainer(model=model,
                           tokenizer=tokenizer,
                           args=training_args,
                           **data_module)

    if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")):
        trainer.train(resume_from_checkpoint=True)
    else:
        trainer.train()
    trainer.save_state()

    model.config.use_cache = True

    if training_args.lora_enable:
        state_dict = get_peft_state_maybe_zero_3(
            model.named_parameters(), training_args.lora_bias
        )
        non_lora_state_dict = get_peft_state_non_lora_maybe_zero_3(
            model.named_parameters()
        )
        if training_args.local_rank == 0 or training_args.local_rank == -1:
            model.config.save_pretrained(training_args.output_dir)
            model.save_pretrained(training_args.output_dir, state_dict=state_dict)
            torch.save(non_lora_state_dict, os.path.join(training_args.output_dir, 'non_lora_trainables.bin'))
    else:
        safe_save_model_for_hf_trainer(trainer=trainer,
                                       output_dir=training_args.output_dir)


if __name__ == '__main__':
    train(attn_implementation="flash_attention_2")