大模型lora微调-chatglm2

通义千问大模型微调源码（chatglm2 微调失败，训练通义千问成功）：GitHub - hiyouga/LLaMA-Efficient-Tuning: Easy-to-use LLM fine-tuning framework (LLaMA-2, BLOOM, Falcon, Baichuan, Qwen, ChatGLM2)Easy-to-use LLM fine-tuning framework (LLaMA-2, BLOOM, Falcon, Baichuan, Qwen, ChatGLM2) - GitHub - hiyouga/LLaMA-Efficient-Tuning: Easy-to-use LLM fine-tuning framework (LLaMA-2, BLOOM, Falcon, Baichuan, Qwen, ChatGLM2)https://github.com/hiyouga/LLaMA-Efficient-Tuning

训练数据准备，进入LLaMA-Efficient-Tuning根目录后，有一个data目录，

创建一个自己的训练数据文件，录入内容格式如下：

[{"instruction":"阅读下列短文，从每题所给的四个选项《A、 B、 C和D)中。选出最佳选项。",

"input":"sdfdsg",

"output:"A"}]训练数据文件配置到dataset_info.json中

训练命令：

vim run.sh

Model_path="model_path"
save_Model_dir="save_model"
dataset="qwentest"

CUDA_VISIBLE_DEVICES=0

扫描二维码关注公众号，回复： 16900707 查看本文章

torchrun --nproc_per_node 1 src/train_bash.py \
--model_name_or_path ${Model_path} \
--stage sft \

--do_train \
--dataset ${dataset} \
--dataset_dir data \
--finetuning_type lora \
--output_dir ${save_Model_dir} \
--overwrite_cache \
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 1 \
--lr_scheduler_type cosine \
--logging_steps 10 \
--save_steps 1000 \

--learning_rate 1e-3 \
--num_train_epochs 10.0 \
--plot_loss \

--lora_target c_attn \
--template chatml \
--fp16

执行训练：

3、执行nohup ./finetune_llm.sh &

4、训练完的模型调用：

from transformers import AutoTokenizer,AutoModelForCausalLM, AutoConfig
from peft import PeftModel

from transformers.generation import GenerationConfig
import os

model_path = "原始模型"
ckpt_path = "lora微调后的模型"

save_pre_name = os.path.basename(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

base_model = AutoModelForCausaLLM.from_pretrained(model_path,device_map={"":0},trust_remote_code=True)

base_model.generation_config = GenerationConfig.from_pretrained(model_path, trust_remote_code=True)

lora_model = PeftModel.from_pretrained(base_model, ckpt_path)

lora_model.to('cuda:4')
lora_model.eval()

input = ""
line_format=input.replace("######","\n")

llm_question="ddsfdsfddsfsdg\nA. 非常满意\nB. 满意\nC. 不满意\nD. 非常不满意\n答案：".format(line_format)

inputs = tokenizer(llm_question, return_tensors='pt')

inputs = inputs.to(lora_model.device)

pred = lora_model.generate(**inputs, max_new_tokens=500)

predict_txt = tokenizer.decode(pred.cpu()[0], skip_special_tokens=True).replace("\n", " ")

print(predict_txt)

关于chatglm2模型微调：（chatglm2成功）： https://github.com/hiyouga/ChatGLM-Efficient-Tuning