HiFT 0.0.5

PyTorch implementation of 'HiFT: A Hierarchical Full Parameter Fine-Tuning Strategy', a memory-efficient approach to adapt a large pre-trained deep learning model.

HiFT: A Hierarchical Full Parameter Fine-Tuning Strategy

This repo contains the source code of the Python package HiFT and several examples of how to integrate it with PyTorch models, such as those in Hugging Face. We only support PyTorch for now. See our paper for a detailed description of ·HiFT. HiFT supports FPFT of 7B models for 24G GPU memory devices under mixed precision without using any memory saving techniques and various optimizers including AdamW, AdaGrad, SGD, etc.

HiFT: A Hierarchical Full Parameter Fine-Tuning Strategy
Yongkang Liu, Yiqun Zhang, Qian Li, Tong Liu, Shi Feng, Daling Wang, Yifei Zhang, Hinrich Schütze
Paper: https://arxiv.org/abs/2401.15207

News

26/1/2024: Publish the first version of HiFT manuscript
25/2/2024: Publish the second version of HiFT manuscript and source code
1/5/2024: Updated HiFT support for LoRA

Repository Overview

There are several directories in this repo:

• hift/ contains the source code for the package hift, which needs to be installed to run the examples we provide;
• examples contains an example implementation of HiFT in BERT, RoBERTa, GPT-2, GPT-Neo,GPT-NeoX,OPT and LLaMA using our package.
• scripts contains the script for running examples we provide.
• dsconfig contains configuration files required for mixed precision.

Quickstart

1. Installing hift is simply

pip install hift
# Alternatively
# pip install git+https://github.com/misonsky/HiFT 

2. Import hift package

from hift import HiFTrainer,GetCallBack

3. Add HiFT configuration

@dataclass
class HiFTArguments(ModelArguments):
    HiTaskType: str = field(
        default="SEQ_CLS",
        metadata={"help": ("HiTaskType should be consistent with PEFT TaskType" )},
    )
    group_element: int = field(
        default=1,
        metadata={"help": ("number element for each group parameters" )},
    )
    optimizer_strategy: str = field(
        default="down2up",
        metadata={"help": ("optimizer strategy of ['down2up','down2up','random']" )},
    )
    hier_tuning: bool = field(
        default=False,
        metadata={
            "help": (
                "hierarchical optimization for LLMS"
            )
        },
    )
    lora_tuning:bool = field(
        default=False,
        metadata={
            "help": (
                "whether using lora tuning"
            )
        },
    )
    freeze_layers: List[str] = field(
        default_factory=list,
        metadata={
            "help": (
                "Index of the frozen layer"
            )
        },
    )

parameter introduction

HiTaskType should be consistent with PEFT TaskType.

sequence classification, multiple choice tasks: TaskType.SEQ_CLS

question answering task: TaskType.QUESTION_ANS

sequence labeling task: TaskType.TOKEN_CLS

generation task: TaskType.CAUSAL_LM

group_element: the number of layers included in a block. Default value is ok.

lora_tuning: HiFT fine-tuning in LoRA mode.

freeze_layers: Layers you want to freeze during fine-tuning. You should provide the index of the corresponding layer. The index of the embedding layer is 0, the index of the first layer is 1,...

4. Using HiFT Trainer

HiFT inherits the trainer of huggingface, so you can directly use the trainer provided by hift to replace the original trainer.

if model_args.hier_tuning:#hier_tuning
     trainer = HiFTrainer(
            hiFThandler = GetCallBack(model_args.model_name_or_path),
            HiTaskType = model_args.HiTaskType,
            group_element = model_args.group_element,
            strategy = model_args.optimizer_strategy,
            hier_tuning= model_args.hier_tuning,
            lora_tuning = model_args.lora_tuning,
            freeze_layers = model_args.freeze_layers,
            train_dataset=train_dataset if training_args.do_train else None,
            eval_dataset=eval_dataset if training_args.do_eval else None,
            model=model,
            tokenizer=tokenizer,
            compute_metrics=compute_metrics,
            data_collator=data_collator,
            args=training_args
        )
else:
    trainer = Trainer(
            model=model,
            args=training_args,
            train_dataset=train_dataset if training_args.do_train else None,
            eval_dataset=eval_dataset if training_args.do_eval else None,
            compute_metrics=compute_metrics,
            tokenizer=tokenizer,
            data_collator=data_collator,
        )

Register Your Model

Theoretically HiFT supports any model. For a new model:

1. provide the task interface provided by your model in TaskTInterface.
2. For different task types, please provide irregular layer regular expressions for identifying them, such as functions SequenceClassificationSpecialLayer, QuestionAnsweringSpecialLayer, CausalLMSpecialLayer
3. Match different task types with corresponding regular expressions, such as GetSpecialLayer
4. provide a regular expression in pattern_name that identifies all layers.
5. The process of extracting identifiers for each layer is provided in group_model.****

class ModelCallBack(HiFTCallBack):
    TaskTInterface = [TaskType.SEQ_CLS,TaskType.QUESTION_ANS,TaskType.CAUSAL_LM]
    def __init__(self,freeze_layers,strategy,lora_tuning=False):
        super().__init__(freeze_layers,strategy)
        self.number_position = 5 if lora_tuning else 3
    @classmethod
    def SequenceClassificationSpecialLayer(cls):
       special_layers = [r"xxxx","xxxx","xxxx"]
       return special_layers
    @classmethod
    def QuestionAnsweringSpecialLayer(cls):
        special_layers = [r"xxxx","xxxx","xxxx"]
        return special_layers
    @classmethod
    def CausalLMSpecialLayer(cls):
        special_layers = [r"xxxx","xxxx","xxxx"]
        return special_layers
    @classmethod
    def GetSpecialLayer(cls,taskType):
        logger.warning("For OPT the HiTaskType should be {}".format(" , ".join(cls.TaskTInterface)))
        assert taskType in cls.TaskTInterface
        if taskType == TaskType.SEQ_CLS:
             return cls.SequenceClassificationSpecialLayer()
        if taskType == TaskType.TOKEN_CLS:
             return cls.TokenClassificationSpecialLayer()
        if taskType == TaskType.CAUSAL_LM:
            return cls.CausalLMSpecialLayer()
    def pattern_name(self,special_layers):
        patterns = [rf'\.\d+\.']
        patterns.extend([rf'{layer}' for layer in special_layers])
        pattern = '|'.join(patterns)
        return pattern
    def check_selection(self,elements,name_search):
        pattern_element = ["\."+element+"\." if element.isdigit() else element for element in elements]
        if len(name_search) >1:
            name_search = name_search[:1]
        assert len(name_search)==1
        signal_value = [1 if len(re.compile(element).findall(name_search[0]))>0 else 0 for element in pattern_element]
        if sum(signal_value)<=0:
            return False
        else:
            return True
    def group_model(self,model,special_layers,num_position=2,lora_tuning=False):
        ....
        

Introduction

The detailed training process is shown in Algorithm. The first step is to determine the update strategy. Then freeze all layers. The layers to be updated, denoted by $E$, are selected from the queue $Q$ based on the parameter $m$. The selected layer $E$ is removed from head of the queue $Q$ and added to the tail of $Q$ to wait for the next update. Select the parameter $\theta_s$ that needs to be updated from $M$ based on $E$, set the parameter $\theta_s$ to a computable gradient state and set the update parameter group of optimizer $P$ to $\theta_s$. Before parameter updates, the states parameters of optimizer $P$ related to $\theta_s$ could be moved to GPU devices. After the completion of weight updates, the corresponding gradients are clean up and optimizer states parameters are moved to CPU. When all layers have been updated once, adjust the learning rate once.

HiFT iteratively updates a subset of parameters at each training step, and it will modify the full parameter after multiple steps. This vastly reduces the GPU memory requirements for fine-tuninglarge language models enables efficient task-switching during deployment all without introducing inference latency. HiFT also outperforms several other adaptation methods including adapter, prefix-tuning, and fine-tuning.

HiFT is a model-independent and optimizer-independent full-parameter fine-tuning method that can be integrated with the PEFT method.

optimizers: The latest version of HiFT is adapted to the Adam, AdamW, SGD, Adafactor and Adagrad optimizers.

Model: The latest version of HiFT supports BERT, RoBERTa, GPT-2, GPTNeo,GPT-NeoX,OPT and LLaMA-based models.

Experiments on OPT-13B (with 1000 examples). ICL: in-context learning; LP: linear probing; FPFT: full fine-tuning; Prefix: prefix-tuning. All experiments use prompts from MeZO.

GPU memory usage of fine-tuning LLaMA (7B) on the E2E dataset. Total represents the total memory used during fine-tuning. Mixed represents fine-tuning with standard mixed precision and Mixed^Hi^ represents the mixed precision adapted to HiFT. Para represents the memory occupied by the model parameters; Gra represents the memory occupied by the gradient; Sta represents the memory occupied by the optimizer state. PGS represents the sum of memory occupied by parameters , gradients and optimizer state .

Mixed Precision

Source Code

class FP16_Optimizer(DeepSpeedOptimizer):
    def __init__(self,
       init_optimizer,
       deepspeed=None,
       static_loss_scale=1.0,
       dynamic_loss_scale=False,
       initial_dynamic_scale=2**32,
       dynamic_loss_args=None,
       verbose=True,
       mpu=None,
       clip_grad=0.0,
       fused_adam_legacy=False,
       has_moe_layers=False,
       timers=None):
                 
       ....
       self.fp16_groups = []
       self.fp16_groups_flat = []
       self.fp32_groups_flat = []
                 
       ...
                 
       for i, param_group in enumerate(self.optimizer.param_groups):
           ...
           self.fp32_groups_flat.append(self.fp16_groups_flat[i].clone().float().detach())
           ...
                            

The memory required to load 1B parameters is 3.72GB (10^9 $\times$ 4 /1024/1024/1024). Standard mixed precision stores both single-precision and half-precision model parameters. Assuming you are using standard mixed precision fine-tuning of the 7B model, compared with single-precision fine-tuning, mixed precision requires an additional about 13G GPU memory overhead to store half-precision model parameters. Only when the dynamic GPU memory reduction reaches 13GB does mixed precision demonstrate its advantages. This requires using large batch size.

We reimplement the mixed-precision algorithm to adapt to HiFT's fine-tuning algorithm, which ensures that single-precision model parameters do not incur additional GPU memory overhead.

Citation

@article{liu2024hift,
  title={HiFT: A Hierarchical Full Parameter Fine-Tuning Strategy},
  author={Liu, Yongkang and Zhang, Yiqun and Li, Qian and Feng, Shi and Wang, Daling and Zhang, Yifei and Sch{\"u}tze, Hinrich},
  journal={arXiv preprint arXiv:2401.15207},
  year={2024}
}

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution.