|
============ |
|
修改训练配置 |
|
============ |
|
|
|
XTuner 的训练由 MMEngine |
|
的训练器提供支持,用户可以通过修改配置文件(config)中的特定参数,来修改对应的训练配置。以 |
|
`internlm2_chat_7b_qlora_oasst1_e3 <https: |
|
为例,本节将首先速览配置文件中各个参数的含义,之后讲解常见配置的修改方式。 |
|
|
|
配置文件速览 |
|
============ |
|
|
|
XTuner 使用 MMEngine 的「纯 Python 风格的配置文件」,直接利用 ``import`` |
|
机制使用一些类或函数。 |
|
|
|
.. tip:: |
|
|
|
如果您期望深入了解 MMEngine 「纯 Python |
|
风格的配置文件」的特性、优势,请参考 |
|
`这里 <https: |
|
|
|
.. code:: python |
|
|
|
# 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 import process_hf_dataset |
|
from xtuner.dataset.collate_fns import default_collate_fn |
|
from xtuner.dataset.map_fns import oasst1_map_fn, template_map_fn_factory |
|
from xtuner.engine.hooks import (DatasetInfoHook, EvaluateChatHook, |
|
VarlenAttnArgsToMessageHubHook) |
|
from xtuner.engine.runner import TrainLoop |
|
from xtuner.model import SupervisedFinetune |
|
from xtuner.utils import PROMPT_TEMPLATE |
|
|
|
####################################################################### |
|
# PART 1 Settings # |
|
####################################################################### |
|
# Model |
|
pretrained_model_name_or_path = 'internlm/internlm2-chat-7b' # 设置 LLM 路径或 HuggingFace Hub ID |
|
use_varlen_attn = False # 是否使用 varlen_attention |
|
|
|
# Data |
|
data_path = 'timdettmers/openassistant-guanaco' # 设置 dataset 路径或 HuggingFace Hub ID,以用于 datasets.load_dataset |
|
prompt_template = PROMPT_TEMPLATE.internlm2_chat # 设置对话模版 |
|
max_length = 2048 # 设置训练数据截断长度 |
|
pack_to_max_length = True # 是否将多条样本打包为一条最长长度的样本 |
|
|
|
# Scheduler & Optimizer |
|
batch_size = 1 # per_device # 每个设备的样本个数 |
|
accumulative_counts = 16 # 梯度累计数 |
|
dataloader_num_workers = 0 # dataloader worker 数 |
|
max_epochs = 3 # 训练迭代代数 |
|
optim_type = AdamW # 优化器 |
|
lr = 2e-4 # 学习率 |
|
betas = (0.9, 0.999) # AdamW 优化器 betas |
|
weight_decay = 0 # AdamW 优化器权重衰减 |
|
max_norm = 1 # grad clip # 梯度裁剪 |
|
warmup_ratio = 0.03 # warmup 比率 |
|
|
|
# Save |
|
save_steps = 500 # checkpoint 保存间隔(iter 数) |
|
save_total_limit = 2 # 最大保存 checkpoint 个数,-1 表示无限制 |
|
|
|
# Evaluate the generation performance during the training |
|
evaluation_freq = 500 # 验证对话效果的执行间隔(iter 数) |
|
SYSTEM = '' # 验证对话效果的 system 字段 |
|
evaluation_inputs = [ # 验证对话效果时的测试问题 |
|
'请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai' |
|
] |
|
|
|
####################################################################### |
|
# PART 2 Model & Tokenizer # |
|
####################################################################### |
|
tokenizer = dict( # 构建 tokenizer |
|
type=AutoTokenizer.from_pretrained, |
|
pretrained_model_name_or_path=pretrained_model_name_or_path, |
|
trust_remote_code=True, |
|
padding_side='right') |
|
|
|
model = dict( # 构建 model |
|
type=SupervisedFinetune, |
|
use_varlen_attn=use_varlen_attn, |
|
llm=dict( # 构建 LLM |
|
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( # 量化配置(保留则为 4 比特,删除则为正常浮点) |
|
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( # LoRA 配置(保留则使用 LoRA 微调,删除则使用全量微调) |
|
type=LoraConfig, |
|
r=64, |
|
lora_alpha=16, |
|
lora_dropout=0.1, |
|
bias='none', |
|
task_type='CAUSAL_LM')) |
|
|
|
####################################################################### |
|
# PART 3 Dataset & Dataloader # |
|
####################################################################### |
|
train_dataset = dict( # 构建训练数据集 |
|
type=process_hf_dataset, |
|
dataset=dict(type=load_dataset, path=data_path), # 调用 datasets.load_dataset 接口 |
|
tokenizer=tokenizer, |
|
max_length=max_length, |
|
dataset_map_fn=oasst1_map_fn, # 选择匹配的数据集 map_fn |
|
template_map_fn=dict( |
|
type=template_map_fn_factory, template=prompt_template), |
|
remove_unused_columns=True, |
|
shuffle_before_pack=True, |
|
pack_to_max_length=pack_to_max_length, |
|
use_varlen_attn=use_varlen_attn) |
|
|
|
train_dataloader = dict( # 构建训练数据集的 DataLoader |
|
batch_size=batch_size, |
|
num_workers=dataloader_num_workers, |
|
dataset=train_dataset, |
|
sampler=dict(type=DefaultSampler, shuffle=True), |
|
collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn)) |
|
|
|
####################################################################### |
|
# PART 4 Scheduler & Optimizer # |
|
####################################################################### |
|
# optimizer |
|
optim_wrapper = dict( # 构建优化器 |
|
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='float16') |
|
|
|
# learning policy |
|
# More information: https: |
|
param_scheduler = [ # 设置学习率 scheduler |
|
dict( |
|
type=LinearLR, # warmup 阶段 |
|
start_factor=1e-5, |
|
by_epoch=True, |
|
begin=0, |
|
end=warmup_ratio * max_epochs, |
|
convert_to_iter_based=True), |
|
dict( |
|
type=CosineAnnealingLR, # Cosine 学习率衰减阶段 |
|
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) # 设置训练迭代代数 |
|
|
|
####################################################################### |
|
# PART 5 Runtime # |
|
####################################################################### |
|
# Log the dialogue periodically during the training process, optional |
|
custom_hooks = [ # 定义 Hooks |
|
dict(type=DatasetInfoHook, tokenizer=tokenizer), # 在训练前打印可视化打印数据样本 |
|
dict( |
|
type=EvaluateChatHook, # 在训练时测试对话效果 |
|
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)] # vallen_attention 依赖的 Hook |
|
|
|
# 以下均为默认配置,如需调整请参考 MMEngine 文档及代码 |
|
|
|
# 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=None, deterministic=False) |
|
|
|
# set log processor |
|
log_processor = dict(by_epoch=False) |
|
|
|
常见训练配置修改 |
|
======================= |
|
|
|
模型 |
|
------------ |
|
|
|
使用其他 LLM 模型? |
|
~~~~~~~~~~~~~~~~~~~~~~~~ |
|
1. 修改 ``pretrained_model_name_or_path``\ ,其将应用至 ``model.llm`` 和 ``tokenizer`` 的初始化中。 |
|
#. 修改 ``prompt_template`` 以适配所选择的 LLM。 |
|
|
|
使用 ModelScope 模型? |
|
~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
|
1. 参考 `文档 <../preparation/pretrained_model.md>`__ 将其下载至本地 |
|
2. 修改\ ``pretrained_model_name_or_path``\ 。 |
|
|
|
使用 openMind 模型? |
|
~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
|
可在配置文件中新增 ``model_resource`` 参数, ``args`` 用作可变参数(如下载私有模型需传入token的情况): |
|
|
|
.. code:: python |
|
from openmind_hub import snapshot_download |
|
|
|
# Model |
|
pretrained_model_name_or_path = 'Tianjin_Ascend/Qwen1.5-4B' |
|
model_resource = { |
|
"fn": snapshot_download, |
|
"args":{ |
|
# "token":"xxxxxxxxxx" |
|
} |
|
} |
|
|
|
微调类型 |
|
------------- |
|
|
|
.. tip:: |
|
XTuner 内置的配置文件以 QLoRA 微调为主,但并不意味着 XTuner 仅支持 QLoRA |
|
微调。用户可以通过修改配置文件中的 ``model`` 来决定微调类型。 |
|
|
|
|
|
QLoRA 微调 |
|
~~~~~~~~~~~~~~~~~ |
|
|
|
.. code:: python |
|
|
|
model = dict( |
|
...... |
|
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'), |
|
......) |
|
|
|
|
|
LoRA 微调 |
|
~~~~~~~~~~~~~~~~ |
|
|
|
.. tip:: |
|
|
|
在 QLoRA 设置的基础上,将 `quantization_config` 设置为 None,就切换成了 LoRA 微调 |
|
|
|
.. code:: python |
|
|
|
model = dict( |
|
...... |
|
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=None), |
|
lora=dict( |
|
type=LoraConfig, |
|
r=64, |
|
lora_alpha=16, |
|
lora_dropout=0.1, |
|
bias='none', |
|
task_type='CAUSAL_LM'), |
|
......) |
|
|
|
|
|
全参数微调 |
|
~~~~~~~~~~~~~~~~~~ |
|
.. tip:: |
|
|
|
将 `lora` 和 `quantization_config` 都设置为 None,就切换到了全参数训练模式 |
|
|
|
.. code:: python |
|
|
|
model = dict( |
|
...... |
|
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=None), |
|
lora=None, |
|
......) |
|
|
|
|
|
|
|
|
|
数据集 |
|
-------------- |
|
|
|
请参考 `训练` 章节文档。 |
|
|
|
优化器 |
|
----------- |
|
|
|
使用其他优化器? |
|
~~~~~~~~~~~~~~~~~~~~ |
|
|
|
- 方法 1:修改 ``optim_type``\ (例如 ``optim_type=torch.optim.SGD``\ ),其将应用至 ``optim_wrapper.optimzer``\ 。 |
|
- 方法 2:忽略 ``optim_type``\ ,直接修改 ``optim_wrapper.optimzer``\ 。 |
|
|
|
|
|
修改优化器参数配置? |
|
~~~~~~~~~~~~~~~~~~~~~~~~ |
|
|
|
- 方法 1:修改 ``lr``\ 、\ ``weight_decay`` 等参数,其将应用至 ``optim_wrapper.optimzer``\ 。 |
|
- 方法 2:直接修改 ``optim_wrapper.optimzer``\ 。 |
|
|
|
迭代次数 |
|
--------------- |
|
|
|
调整迭代次数? |
|
~~~~~~~~~~~~~~~~~~~~~ |
|
|
|
- 修改 ``max_epochs`` 参数。 |
|
|
|
保存 Checkpoint 间隔 |
|
--------------------------- |
|
|
|
调整保存间隔? |
|
~~~~~~~~~~~~~~~~~~~~~ |
|
|
|
- 修改 ``save_steps`` 参数。 |
|
|
|
调整最大保存 checkpoint 个数? |
|
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
|
|
|
- 修改 ``save_total_limit`` 参数。 |
|
|
|
训练间对话评测 |
|
---------------------- |
|
|
|
调整对话评测间隔? |
|
~~~~~~~~~~~~~~~~~~~~~~~~~~ |
|
|
|
- 修改 ``evaluation_freq`` 参数。 |
|
|
|
调整对话评测的 system 字段? |
|
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
|
|
|
- 修改 ``SYSTEM`` 参数。 |
|
|
|
调整对话评测的测试指令? |
|
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
|
|
|
- 修改 ``evaluation_inputs`` 参数。 |
|
|
|
GPU 数 |
|
-------------- |
|
|
|
XTuner |
|
的多卡训练由启动命令决定,而非配置文件。用户可以参考下列命令启动多卡训练: |
|
|
|
.. code:: bash |
|
|
|
# 单卡 |
|
xtuner train ${CONFIG} |
|
# 多卡 |
|
(DIST) NPROC_PER_NODE=${GPU_NUM} xtuner train ${CONFIG} |
|
(SLURM) srun ${SRUN_ARGS} xtuner train ${CONFIG} --launcher slurm |
|
|
|
DeepSpeed |
|
------------------ |
|
|
|
XTuner 的 DeepSpeed |
|
优化由启动命令决定,而非配置文件。用户可以参考下列命令启用 DeepSpeed |
|
优化: |
|
|
|
.. code:: bash |
|
|
|
xtuner train ${CONFIG} --deepspeed ${DS_CONFIG} |
|
|
|
.. note:: |
|
|
|
XTuner 内置了多个 DeepSpeed 配置文件(即命令中的 |
|
``${DS_CONFIG}``\ ),用户可以直接使用,具体文件见 |
|
`这里 <https: |
|
|
|
.. code:: bash |
|
|
|
xtuner train ${CONFIG} --deepspeed [deepspeed_zero1,deepspeed_zero2,deepspeed_zero2_offload,deepspeed_zero3,deepspeed_zero3_offload] |
|
|
|
.. note:: |
|
部分参数会在 DeepSpeed Config 和 XTuner Config 中重复定义(例如 batch |
|
size等)。此时相关配置会以 XTuner Config 为准: |
|
|
|
- ``gradient_accumulation_steps`` 会被 XTuner Config 中的 |
|
``accumulative_counts`` 设置覆盖。 |
|
|
|
- ``train_micro_batch_size_per_gpu`` 会被 XTuner Config 中的 |
|
``train_dataloader.batch_size`` 设置覆盖。 |
|
|
|
- ``gradient_clipping`` 会被 XTuner Config 中的 |
|
``optim_wrapper.clip_grad.max_norm`` 设置覆盖。 |
|
|
|
- XTuner 会根据所使用的 GPU 架构自动选择 ``fp16`` 或 ``bf16`` 训练。 |
|
|
|
其他 |
|
---------- |
|
|
|
如有遗漏或特定需求,欢迎提出 |
|
`issue <https: |
|
|