InternLM
pack_to_max_length = False后卡死
pack_to_max_length = True 可以正常运行，pack_to_max_length = False后卡死，log打到"Checkpoints will be saved to...”就卡住了，也不报错，请问是怎么回事啊？
很抱歉给您的使用带来不便。为了我们进一步定位问题，麻烦提供一下训练 config 以及启动脚本
十分感谢回复！下面是config、启动脚本和环境：
config：
Copyright (c) OpenMMLab. All rights reserved.
"""Data format:
[{ "messages": [ { "role": "system", "content": "xxx." }, { "role": "user", "content": "xxx." }, { "role": "assistant", "content": "xxx.", "loss": false}, { "role": "user", "content": "xxx." }, { "role": "assistant", "content": "xxx.", "loss": true} ] }, ... ] """ # noqa: E501 import torch from datasets import load_dataset from mmengine.config import read_base from mmengine.dataset import DefaultSampler from xtuner.dataset.samplers import LengthGroupedSampler 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 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
with read_base(): from .map_f import single_turn_map_f as dataset_map_fn
#######################################################################
PART 1 Settings
#######################################################################
Model
pretrained_model_name_or_path = "/home/model/Qwen2-7B"
use_varlen_attn = False
Data
data_files = ["/home/data/train.json"]
prompt_template = PROMPT_TEMPLATE.qwen_chat
max_length = 100000
pack_to_max_length = False
Scheduler & Optimizer
batch_size = 1 # per_device accumulative_counts = 6 # bs = 1 GPU * 1 batch_size_per_device * 16 acc dataloader_num_workers = 0 max_epochs = 10 optim_type = AdamW lr = 5e-6 betas = (0.9, 0.999) weight_decay = 0 max_norm = 1 # grad clip warmup_ratio = 0.03
Save
save_steps = 100 save_total_limit = -1 # Maximum checkpoints to keep (-1 means unlimited)
Evaluate the generation performance during the training
evaluation_freq = 500 SYSTEM = '' evaluation_inputs = [ '请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai' ]
sequence_parallel_size = 8
#######################################################################
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', eos_token='<|im_end|>')
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, torch_dtype=torch.float16))
#######################################################################
PART 3 Dataset & Dataloader
####################################################################### train_dataset = dict( type=process_hf_dataset, dataset=dict(type=load_dataset, path='json', data_files=data_files), tokenizer=tokenizer, max_length=max_length, dataset_map_fn=dataset_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( 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( 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, 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=None, deterministic=False)
set log processor
log_processor = dict(by_epoch=False)
启动脚本：
GPU_NUM=8 CONFIG_PATH="/home/code/qwen_7b_custom_sft_copy.py" OUT_DIR=/home/result/Qwen2-sft NPROC_PER_NODE=${GPU_NUM} xtuner train ${CONFIG_PATH} --work-dir ${OUT_DIR} --deepspeed deepspeed_zero2
环境： torch==2.1.2 cuda==12.1 transfomers==4.41.2 flash-attn==2.5.9post1 xtuner==0.1.21 deepspeed==0.14.0
@bo-jpg 您这里开了序列并行，需要对应修改sampler为 SequenceParallelSampler ，类似这样 https://github.com/InternLM/xtuner/blob/main/xtuner/configs/llama/llama3_8b/llama3_8b_full_alpaca_e3.py#L93C11-L94
@bo-jpg 您这里开了序列并行，需要对应修改sampler为 SequenceParallelSampler ，类似这样 https://github.com/InternLM/xtuner/blob/main/xtuner/configs/llama/llama3_8b/llama3_8b_full_alpaca_e3.py#L93C11-L94
使用SequenceParallelSampler 可以了，感谢！另外可以顺带解释一下SequenceParallelSampler 的采样方式吗，比起DefaultSampler和LengthGroupedSampler有什么不同？
开启了序列并行之后，dp size = world size // sp size，相同dp_rank不同sp_rank的gpu，需要采样相同的数据，forward前需要对数据沿着seq_len维度做切分。
@HIT-cwh 了解了，感谢！
另外，我的配置： use_varlen_attn = False pack_to_max_length = False batch_size = 1 # per_device accumulative_counts = 6 # bs = 1 GPU * 1 batch_size_per_device * 16 acc max_epochs = 10 我的训练数据有3551条，但是log打印： 07/22 14:34:08 - mmengine - INFO - Num train samples 3427，以及 07/22 14:34:42 - mmengine - INFO - Iter(train) [ 10/34270] lr: 4.3867e-08 eta: 1 day, 4:36:17 time: 3.0058 data_time: 0.0179 memory: 36687 loss: 0.7363 07/22 14:35:12 - mmengine - INFO - Iter(train) [ 20/34270] lr: 9.2552e-08 eta: 1 day, 4:31:40 time: 2.9913 data_time: 0.0314 memory: 49394 loss: 0.6501
请问：1.3551为什么变成了3427？2.为什么总iter数是34270，怎么算的？ 辛苦解答！