环境
微调框架:LLaMA-Efficient-Tuning
训练机器:4*RTX3090TI (24G显存)
python环境:python3.8, 安装requirements.txt
依赖包
一、Lora微调
1、准备数据集
2、训练及测试
1)创建模型输出目录
mkdir -p models/llama2_7b_chat/llama-main/train_models/llama2_7b_chat_muti_gpus_01_epoch10/train_model
2)创建deepspeed配置文件目录
mkdir -p models/baichuan2_13b_chat/deepspeed_config
3)创建deepspeed配置文件
vi models/llama2_7b_chat/llama-main/deepspeed_config/llama2_7b_chat_muti_gpus_01_epoch10.json
{
"bf16": {
"enabled": true
},
"fp16": {
"enabled": "auto",
"loss_scale": 0,
"loss_scale_window": 1000,
"initial_scale_power": 16,
"hysteresis": 2,
"min_loss_scale": 1
},
"optimizer": {
"type": "AdamW",
"params": {
"lr": "auto",
"betas": "auto",
"eps": "auto",
"weight_decay": "auto"
}
},
"scheduler": {
"type": "WarmupDecayLR",
"params": {
"last_batch_iteration": -1,
"total_num_steps": "auto",
"warmup_min_lr": "auto",
"warmup_max_lr": "auto",
"warmup_num_steps": "auto"
}
},
"zero_optimization": {
"stage": 3,
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
},
"offload_param": {
"device": "cpu",
"pin_memory": true
},
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 1e9,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 2e9,
"stage3_max_reuse_distance": 2e9,
"stage3_gather_16bit_weights_on_model_save": true
},
"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"steps_per_print": 2000,
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"wall_clock_breakdown": false
}
4)训练模型
deepspeed --num_gpus 2 --master_port=9902 src/train_bash1.py \
--stage sft \
--model_name_or_path models/llama2_7b_chat/origin_model/Llama-2-7b-chat-hf \
--do_train \
--dataset example1 \
--template llama2 \
--finetuning_type lora \
--lora_target q_proj,v_proj \
--output_dir models/llama2_7b_chat/llama-main/train_models/llama2_7b_chat_muti_gpus_01_epoch10/train_model \
--overwrite_cache \
--per_device_train_batch_size 2 \
--gradient_accumulation_steps 2 \
--lr_scheduler_type cosine \
--logging_steps 10 \
--save_steps 500 \
--learning_rate 5e-5 \
--num_train_epochs 100.0 \
--plot_loss \
--bf16 \
--deepspeed models/llama2_7b_chat/llama-main/deepspeed_config/llama2_7b_chat_muti_gpus_01_epoch10.json
- 测试模型
python src/cli_demo.py \
--model_name_or_path models/llama2_7b_chat/origin_model \
--template baichuan2 \
--finetuning_type lora \
--checkpoint_dir models/llama2_7b_chat/llama-main/train_models/llama2_7b_chat_muti_gpus_01_epoch10/train_model
6)启动服务
python src/web_demo1.py \
--model_name_or_path models/llama2_7b_chat/origin_model/Llama-2-7b-chat-hf \
--template llama2 \
--finetuning_type lora \
--checkpoint_dir models/llama2_7b_chat/llama-main/train_models/llama2_7b_chat_muti_gpus_01_epoch10/train_model