Yi-34B模型使用双卡deepspeed zero2 训练加载模型时占用CPU 内存>200G 不足导致失败

[xxll88]

Reminder

• I have read the README and searched the existing issues.

Reproduction

model

model_name_or_path: /home/ubuntu/Yi-1.5-34B

method

stage: pt
do_train: true
finetuning_type: freeze

template: default

ddp

ddp_timeout: 180000000
deepspeed: examples/deepspeed/ds_z2_config.json

dataset

dataset: qclound,intlcloud
cutoff_len: 1024
#max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16

output

output_dir: ../saves/Yi-1.5-34B/ptqcloud2
save_total_limit: 1
logging_steps: 20
save_steps: 1000
plot_loss: true
overwrite_output_dir: false

train

per_device_train_batch_size: 16 #16
gradient_accumulation_steps: 1 #2
learning_rate: 0.0001
num_train_epochs: 2.0
lr_scheduler_type: cosine
warmup_steps: 0.1
bf16: true #bf16

eval

val_size: 0.001
per_device_eval_batch_size: 1
evaluation_strategy: steps
eval_steps: 500

examples/deepspeed/ds_z2_config.json
{
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"zero_allow_untested_optimizer": true,
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"bf16": {
"enabled": "auto"
},
"zero_optimization": {
"stage": 2,
"allgather_partitions": true,
"allgather_bucket_size": 1e8,
"overlap_comm": true,
"reduce_scatter": true,
"reduce_bucket_size": 5e8,
"contiguous_gradients": true,
"round_robin_gradients": true
}
}

Expected behavior

System Info

No response

Others

No response

[hiyouga]

更新到最新版代码