xtuner
xtuner copied to clipboard
Published
1 month ago
•
InternLM
InternLM
关于训练的一些疑惑
感谢贵团队的贡献！ 最近在使用xtuner训练，但是遇到了一些问题。 1、一些参数的含义不是很清楚，有没有针对每个参数的说明文件呢？
2、我进行sft，但是跑起来后step和我手动算的对不是，config：
# Copyright (c) OpenMMLab. All rights reserved.
from peft import LoraConfig
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 torch.optim import AdamW
from transformers import AutoModelForCausalLM, AutoTokenizer
from xtuner.dataset import process_hf_dataset
from xtuner.dataset.collate_fns import default_collate_fn
from xtuner.dataset.map_fns import pretrain_map_fn
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, SYSTEM_TEMPLATE
from xtuner.dataset.map_fns import alpaca_map_fn, template_map_fn_factory
#######################################################################
# PART 1 Settings #
#######################################################################
sequence_parallel_size=1
# Model
pretrained_model_name_or_path = '/NVME1/elecLLM/models/qwen/Qwen2-72B-Instruct'
use_varlen_attn = False
# Data
data_path = '/NVME1/projects/xtuner/data/online_query_1000.json'
prompt_template = PROMPT_TEMPLATE.qwen_chat
max_length = 32 * 1024
pack_to_max_length = False
# Scheduler & Optimizer
batch_size = 1 # per_device
accumulative_counts = 16 # bs = 1 GPU * 1 batch_size_per_device * 16 acc
dataloader_num_workers = 0
max_epochs = 1
optim_type = AdamW
lr = 2e-5
betas = (0.9, 0.999)
weight_decay = 0
max_norm = 1 # grad clip
warmup_ratio = 0.03
# Save
save_steps = 500
save_total_limit = 2 # Maximum checkpoints to keep (-1 means unlimited)
# Evaluate the generation performance during the training
evaluation_freq = 500
SYSTEM = SYSTEM_TEMPLATE.alpaca
evaluation_inputs = ['上海是', 'Shanghai is']
#######################################################################
# 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=SupervisedFinetune,
use_varlen_attn=use_varlen_attn,
llm=dict(
type=AutoModelForCausalLM.from_pretrained,
pretrained_model_name_or_path=pretrained_model_name_or_path,
trust_remote_code=True),
lora=dict(
type=LoraConfig,
r=8,
lora_alpha=16,
lora_dropout=0.05,
target_modules=['q_proj', 'down_proj', 'o_proj', 'k_proj', 'v_proj', 'up_proj', 'gate_proj'],
modules_to_save=['model.embed_tokens', 'model.norm', 'model.layers.0.input_layernorm', 'model.layers.0.post_attention_layernorm', 'model.layers.1.input_layernorm', 'model.layers.1.post_attention_layernorm', 'model.layers.2.input_layernorm', 'model.layers.2.post_attention_layernorm', 'model.layers.3.input_layernorm', 'model.layers.3.post_attention_layernorm', 'model.layers.4.input_layernorm', 'model.layers.4.post_attention_layernorm', 'model.layers.5.input_layernorm', 'model.layers.5.post_attention_layernorm', 'model.layers.6.input_layernorm', 'model.layers.6.post_attention_layernorm', 'model.layers.7.input_layernorm', 'model.layers.7.post_attention_layernorm', 'model.layers.8.input_layernorm', 'model.layers.8.post_attention_layernorm', 'model.layers.9.input_layernorm', 'model.layers.9.post_attention_layernorm', 'model.layers.10.input_layernorm', 'model.layers.10.post_attention_layernorm', 'model.layers.11.input_layernorm', 'model.layers.11.post_attention_layernorm', 'model.layers.12.input_layernorm', 'model.layers.12.post_attention_layernorm', 'model.layers.13.input_layernorm', 'model.layers.13.post_attention_layernorm', 'model.layers.14.input_layernorm', 'model.layers.14.post_attention_layernorm', 'model.layers.15.input_layernorm', 'model.layers.15.post_attention_layernorm', 'model.layers.16.input_layernorm', 'model.layers.16.post_attention_layernorm', 'model.layers.17.input_layernorm', 'model.layers.17.post_attention_layernorm', 'model.layers.18.input_layernorm', 'model.layers.18.post_attention_layernorm', 'model.layers.19.input_layernorm', 'model.layers.19.post_attention_layernorm', 'model.layers.20.input_layernorm', 'model.layers.20.post_attention_layernorm', 'model.layers.21.input_layernorm', 'model.layers.21.post_attention_layernorm', 'model.layers.22.input_layernorm', 'model.layers.22.post_attention_layernorm', 'model.layers.23.input_layernorm', 'model.layers.23.post_attention_layernorm', 'model.layers.24.input_layernorm', 'model.layers.24.post_attention_layernorm', 'model.layers.25.input_layernorm', 'model.layers.25.post_attention_layernorm', 'model.layers.26.input_layernorm', 'model.layers.26.post_attention_layernorm', 'model.layers.27.input_layernorm', 'model.layers.27.post_attention_layernorm', 'model.layers.28.input_layernorm', 'model.layers.28.post_attention_layernorm', 'model.layers.29.input_layernorm', 'model.layers.29.post_attention_layernorm', 'model.layers.30.input_layernorm', 'model.layers.30.post_attention_layernorm', 'model.layers.31.input_layernorm', 'model.layers.31.post_attention_layernorm', 'model.layers.32.input_layernorm', 'model.layers.32.post_attention_layernorm', 'model.layers.33.input_layernorm', 'model.layers.33.post_attention_layernorm', 'model.layers.34.input_layernorm', 'model.layers.34.post_attention_layernorm', 'model.layers.35.input_layernorm', 'model.layers.35.post_attention_layernorm', 'model.layers.36.input_layernorm', 'model.layers.36.post_attention_layernorm', 'model.layers.37.input_layernorm', 'model.layers.37.post_attention_layernorm', 'model.layers.38.input_layernorm', 'model.layers.38.post_attention_layernorm', 'model.layers.39.input_layernorm', 'model.layers.39.post_attention_layernorm', 'model.layers.40.input_layernorm', 'model.layers.40.post_attention_layernorm', 'model.layers.41.input_layernorm', 'model.layers.41.post_attention_layernorm', 'model.layers.42.input_layernorm', 'model.layers.42.post_attention_layernorm', 'model.layers.43.input_layernorm', 'model.layers.43.post_attention_layernorm', 'model.layers.44.input_layernorm', 'model.layers.44.post_attention_layernorm', 'model.layers.45.input_layernorm', 'model.layers.45.post_attention_layernorm', 'model.layers.46.input_layernorm', 'model.layers.46.post_attention_layernorm', 'model.layers.47.input_layernorm', 'model.layers.47.post_attention_layernorm', 'model.layers.48.input_layernorm', 'model.layers.48.post_attention_layernorm', 'model.layers.49.input_layernorm', 'model.layers.49.post_attention_layernorm', 'model.layers.50.input_layernorm', 'model.layers.50.post_attention_layernorm', 'model.layers.51.input_layernorm', 'model.layers.51.post_attention_layernorm', 'model.layers.52.input_layernorm', 'model.layers.52.post_attention_layernorm', 'model.layers.53.input_layernorm', 'model.layers.53.post_attention_layernorm', 'model.layers.54.input_layernorm', 'model.layers.54.post_attention_layernorm', 'model.layers.55.input_layernorm', 'model.layers.55.post_attention_layernorm', 'model.layers.56.input_layernorm', 'model.layers.56.post_attention_layernorm', 'model.layers.57.input_layernorm', 'model.layers.57.post_attention_layernorm', 'model.layers.58.input_layernorm', 'model.layers.58.post_attention_layernorm', 'model.layers.59.input_layernorm', 'model.layers.59.post_attention_layernorm', 'model.layers.60.input_layernorm', 'model.layers.60.post_attention_layernorm', 'model.layers.61.input_layernorm', 'model.layers.61.post_attention_layernorm', 'model.layers.62.input_layernorm', 'model.layers.62.post_attention_layernorm', 'model.layers.63.input_layernorm', 'model.layers.63.post_attention_layernorm', 'model.layers.64.input_layernorm', 'model.layers.64.post_attention_layernorm', 'model.layers.65.input_layernorm', 'model.layers.65.post_attention_layernorm', 'model.layers.66.input_layernorm', 'model.layers.66.post_attention_layernorm', 'model.layers.67.input_layernorm', 'model.layers.67.post_attention_layernorm', 'model.layers.68.input_layernorm', 'model.layers.68.post_attention_layernorm', 'model.layers.69.input_layernorm', 'model.layers.69.post_attention_layernorm', 'model.layers.70.input_layernorm', 'model.layers.70.post_attention_layernorm', 'model.layers.71.input_layernorm', 'model.layers.71.post_attention_layernorm', 'model.layers.72.input_layernorm', 'model.layers.72.post_attention_layernorm', 'model.layers.73.input_layernorm', 'model.layers.73.post_attention_layernorm', 'model.layers.74.input_layernorm', 'model.layers.74.post_attention_layernorm', 'model.layers.75.input_layernorm', 'model.layers.75.post_attention_layernorm', 'model.layers.76.input_layernorm', 'model.layers.76.post_attention_layernorm', 'model.layers.77.input_layernorm', 'model.layers.77.post_attention_layernorm', 'model.layers.78.input_layernorm', 'model.layers.78.post_attention_layernorm', 'model.layers.79.input_layernorm', 'model.layers.79.post_attention_layernorm'],
bias='none',
task_type='CAUSAL_LM'))
# model = dict(
# type=SupervisedFinetune,
# use_varlen_attn=use_varlen_attn,
# llm=dict(
# type=AutoModelForCausalLM.from_pretrained,
# pretrained_model_name_or_path=pretrained_model_name_or_path,
# trust_remote_code=True))
#######################################################################
# PART 3 Dataset & Dataloader #
#######################################################################
train_dataset = dict(
type=process_hf_dataset,
# dataset=dict(type=load_dataset, path='json', data_files=data_files),
dataset=dict(
type=load_dataset, path='json', data_files=dict(train=data_path)),
tokenizer=tokenizer,
max_length=max_length,
dataset_map_fn=alpaca_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)
sampler = SequenceParallelSampler \
if sequence_parallel_size > 1 else DefaultSampler
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=default_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='float16')
# 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)
]
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=1),
# 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)
配置文件是参考提供的alpaca sft文件修改的，自定义数据集online_query_1000.json是1000条单轮对话数据，关键参数bs=1，accumulative_counts=16，sequence_parallel_size=1，max_epochs=1，我的理解是，step=1000/16=62.5，但实际有125个step，为什么会出现这种状况，bs和accumulative_counts的实际含义是什么呢？
3、关于截断和拼接的逻辑 使用config文件是跟上一个一样，机子是8*A100（80G），max_length=32k和sequence_parallel_size=1，这个配置我跑pt的时候会OOM，但是跑sft，单卡只占到30G（ds优化策略相同），想问下是因为在pack_to_max_length设为False的情况下，样本的长度不会填充到32k吗？ 如果将pack_to_max_length设为True，一条训练数据中拼接了多个指令，训练时指令间是否能看到彼此，会不会相互影响？
你解决了吗？我也非常困惑... 这个step数，我调整accumulative_counts 对于总step数没有什么影响... 非常奇怪
+1
