llama3_8b_instruct_dpo.py

# Copyright (c) OpenMMLab. All rights reserved.
import torch
from datasets import load_dataset
from mmengine.dataset import DefaultSampler
from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
                            LoggerHook, ParamSchedulerHook)
from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
from peft import LoraConfig
from torch.optim import AdamW
from transformers import (AutoModelForCausalLM, AutoTokenizer,
                          BitsAndBytesConfig)

from xtuner.dataset.collate_fns.preference_collate_fn import \
    preference_collate_fn
from xtuner.dataset.preference_dataset import (build_preference_dataset,
                                               orpo_dpo_mix_40k_map_fn,ultrafeedback_dpo_map_fn,load_jsonl_dataset)
from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook,
                                 VarlenAttnArgsToMessageHubHook)
from xtuner.engine.runner import TrainLoop
from xtuner.model.dpo import DPO
from xtuner.parallel.sequence import SequenceParallelSampler
from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE

#######################################################################
#                          PART 1  Settings                           #
#######################################################################
# Model
pretrained_model_name_or_path = ''
save_steps = 250
lr = 1e-6  # refer to alignment handbook

use_varlen_attn = True
dpo_loss_type = 'sigmoid'  # One of ['sigmoid', 'hinge', 'ipo', 'kto_pair', 'sppo_hard', 'nca_pair', 'robust']  # noqa: E501
loss_beta = 0.1
label_smoothing = 0.0

# Data
# prompt_template = PROMPT_TEMPLATE.llama3_chat
prompt_template = PROMPT_TEMPLATE.llama2_chat # llama2_chat
max_length = 32768
max_packed_length = max_length

# parallel
sequence_parallel_size = 2

# Scheduler & Optimizer
batch_size = 1  # per_device
accumulative_counts = 1
accumulative_counts *= sequence_parallel_size
dataloader_num_workers = 0
max_epochs = 2
optim_type = AdamW
betas = (0.9, 0.999)
weight_decay = 0
max_norm = 1  # grad clip
warmup_ratio = 0.1 # 0.03

# Save
save_total_limit = -1  # Maximum checkpoints to keep (-1 means unlimited)

# Evaluate the generation performance during the training
evaluation_freq = 90
SYSTEM = SYSTEM_TEMPLATE.alpaca
evaluation_inputs = [
    'What famous British author, known for his tales of mystery and the macabre, shares his initials with a common abbreviation for "rest in peace"?',  # noqa: E501
    'Please tell me five scenic spots in Shanghai',
    '890729 - 425663? Only respond with math and no words.'
]

#######################################################################
#                      PART 2  Model & Tokenizer                      #
#######################################################################
tokenizer = dict(
    type=AutoTokenizer.from_pretrained,
    pretrained_model_name_or_path=pretrained_model_name_or_path,
    trust_remote_code=True,
    padding_side='right')

model = dict(
    type=DPO,
    loss_type=dpo_loss_type,
    use_varlen_attn=use_varlen_attn,
    beta=loss_beta,
    label_smoothing=label_smoothing,
    llm=dict(
        type=AutoModelForCausalLM.from_pretrained,
        pretrained_model_name_or_path=pretrained_model_name_or_path,
        trust_remote_code=True,
        torch_dtype=torch.bfloat16,
        ),
    ref_llm=dict(  ##### initialization of ref_llm #######
        type=AutoModelForCausalLM.from_pretrained,
        pretrained_model_name_or_path=pretrained_model_name_or_path,
        trust_remote_code=True,
        torch_dtype=torch.bfloat16,
        ),
)
    # llm=dict(
    #     type=AutoModelForCausalLM.from_pretrained,
    #     pretrained_model_name_or_path=pretrained_model_name_or_path,
    #     trust_remote_code=True,
    #     torch_dtype=torch.float16,
    #     quantization_config=dict(
    #         type=BitsAndBytesConfig,
    #         load_in_4bit=True,
    #         load_in_8bit=False,
    #         llm_int8_threshold=6.0,
    #         llm_int8_has_fp16_weight=False,
    #         bnb_4bit_compute_dtype=torch.float16,
    #         bnb_4bit_use_double_quant=True,
    #         bnb_4bit_quant_type='nf4')),
    # lora=dict(
    #     type=LoraConfig,
    #     r=64,
    #     lora_alpha=16,
    #     lora_dropout=0.1,
    #     bias='none',
    #     task_type='CAUSAL_LM'))
#######################################################################
#                      PART 3  Dataset & Dataloader                   #
#######################################################################
sampler = SequenceParallelSampler \
    if sequence_parallel_size > 1 else DefaultSampler

train_dataset = dict(
    type=build_preference_dataset,
    dataset=dict(type=load_jsonl_dataset,data_files=[""]), 
    # dataset=dict(type=load_dataset, path='llamafactory/ultrafeedback_binarized'), # mlabonne/orpo-dpo-mix-40k
    tokenizer=tokenizer,
    max_length=max_length,
    dataset_map_fn=ultrafeedback_dpo_map_fn,
    is_dpo=True,
    is_reward=False,
    reward_token_id=-1,
    num_proc=32,
    use_varlen_attn=use_varlen_attn,
    max_packed_length=max_packed_length,
    shuffle_before_pack=True,
)

train_dataloader = dict(
    batch_size=batch_size,
    num_workers=dataloader_num_workers,
    dataset=train_dataset,
    sampler=dict(type=sampler, shuffle=True),
    collate_fn=dict(
        type=preference_collate_fn, use_varlen_attn=use_varlen_attn))

#######################################################################
#                    PART 4  Scheduler & Optimizer                    #
#######################################################################
# optimizer
optim_wrapper = dict(
    type=AmpOptimWrapper,
    optimizer=dict(
        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
    accumulative_counts=accumulative_counts,
    loss_scale='dynamic',
    dtype='bfloat16')

# learning policy
# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
param_scheduler = [
    dict(
        type=LinearLR,
        start_factor=1e-5,
        by_epoch=True,
        begin=0,
        end=warmup_ratio * max_epochs,
        convert_to_iter_based=True),
    dict(
        type=CosineAnnealingLR,
        eta_min=0.0,
        by_epoch=True,
        begin=warmup_ratio * max_epochs,
        end=max_epochs,
        convert_to_iter_based=True)
]

# train, val, test setting
train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)

#######################################################################
#                           PART 5  Runtime                           #
#######################################################################
# Log the dialogue periodically during the training process, optional

custom_hooks = [
    dict(type=DatasetInfoHook, tokenizer=tokenizer),
    # dict(
    #     type=EvaluateChatHook,
    #     tokenizer=tokenizer,
    #     every_n_iters=evaluation_freq,
    #     evaluation_inputs=evaluation_inputs,
    #     system=SYSTEM,
    #     prompt_template=prompt_template)
]

if use_varlen_attn:
    custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]

# configure default hooks
default_hooks = dict(
    # record the time of every iteration.
    timer=dict(type=IterTimerHook),
    # print log every 10 iterations.
    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
    # enable the parameter scheduler.
    param_scheduler=dict(type=ParamSchedulerHook),
    # save checkpoint per `save_steps`.
    checkpoint=dict(
        type=CheckpointHook,
        by_epoch=False,
        interval=save_steps,
        max_keep_ckpts=save_total_limit),
    # set sampler seed in distributed evrionment.
    sampler_seed=dict(type=DistSamplerSeedHook),
)

# configure environment
env_cfg = dict(
    # whether to enable cudnn benchmark
    cudnn_benchmark=False,
    # set multi process parameters
    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
    # set distributed parameters
    dist_cfg=dict(backend='nccl'),
)

# set visualizer
visualizer = None

# set log level
log_level = 'INFO'

# load from which checkpoint
load_from = None

# whether to resume training from the loaded checkpoint
resume = False

# Defaults to use random seed and disable `deterministic`
randomness = dict(seed=42, deterministic=False)

# set log processor
log_processor = dict(by_epoch=False)