在昇腾开发环境合并baichuan2-13B模型的lora文件

概述

别人使用lora的方式对baichuan2-13b的模型进行了微调训练，希望我能部署到mindie服务中。

lora微调后生成的文件如下：

# ls checkpoint-lora
adapter_config.json        optimizer.pt  rng_state.pth  special_tokens_map.json   tokenizer_config.json  trainer_state.json
adapter_model.safetensors  README.md     scheduler.pt   tokenization_baichuan.py  tokenizer.model        training_args.bin

# cat checkpoint-lora/adapter_config.json
{
  "alpha_pattern": {},
  "auto_mapping": null,
  "base_model_name_or_path": "/home/xxxxx/baichuan-inc/Baichuan2-13B-Chat", #这里是微调基础模型路径
  "bias": "none",
  "fan_in_fan_out": false,
  "inference_mode": true,
  "init_lora_weights": true,
  "layers_pattern": null,
  "layers_to_transform": null,
  。。。。。。。。。。
}

测试合并前的lora模型

加载合并前的lora模型，需要用AutoPeftModelForCausalLM.from_pretrained生成model实例，其它地方和正常模型都一样。

import torch
import torch_npu
from torch_npu.npu import amp
from torch_npu.contrib import transfer_to_npu
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers.generation.utils import GenerationConfig
from peft import AutoPeftModelForCausalLM

model_path="/home/xxxx/baichuan-inc/Baichuan2-13B-Chat/"
lora_path="/home/xxxx/baichuan-inc/checkpoint-lora/"

tokenizer = AutoTokenizer.from_pretrained(model_path,
    revision="v2.0",
    use_fast=False,
    trust_remote_code=True)
model = AutoPeftModelForCausalLM.from_pretrained(lora_path,
    revision="v2.0",
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True).half().npu().eval()

model.generation_config = GenerationConfig.from_pretrained(model_path, revision="v2.0")

messages = []
messages.append({"role": "user", "content": "讲一个100字左右的故事"})

response = model.chat(tokenizer, messages)
print(response)

将lora微调后的模型合并到baichuan2-13b的基础模型

下面的代码从model_path加载基础模型，从lora_path加载lora模型，最终将合并后的文件保存到merge_path。

import torch
import torch_npu
from torch_npu.npu import amp
from torch_npu.contrib import transfer_to_npu
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers.generation.utils import GenerationConfig
from peft import PeftModel

model_path="/home/xxxx/baichuan-inc/Baichuan2-13B-Chat/"
lora_path="/home/xxxx/baichuan-inc/checkpoint-lora/"
merge_path="/home/xxxx/baichuan-inc/Baichuan2-13B-Chat-lora-merge"


print(f"Loading the Base model from {model_path}")
tokenizer = AutoTokenizer.from_pretrained(model_path,
    revision="v2.0",
    use_fast=False,
    trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(model_path,
    revision="v2.0",
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True)
    #trust_remote_code=True).eval().half().npu()

print(f"Loading the LoRA from {lora_path}")
lora_model = PeftModel.from_pretrained(
        base_model,
        lora_path,
        torch_dtype=torch.float16,
    )


print("Applying the LoRA")
model = lora_model.merge_and_unload()

print(f"Saving the target model to {merge_path}")
model.save_pretrained(merge_path)
print(f"Saving the tokenizer to {merge_path}")
tokenizer.save_pretrained(merge_path)

查看合并后的模型目录：

# ls Baichuan2-13B-Chat-lora-merge/ -lh
total 26G
-rw-r----- 1 root root  780 Jun 24 16:16 config.json
-rw------- 1 root root 1.6K Jun 24 16:16 configuration_baichuan.py
-rw-r----- 1 root root  285 Jun 24 16:16 generation_config.json
-rw------- 1 root root 2.9K Jun 24 16:16 generation_utils.py
-rw-r----- 1 root root 4.6G Jun 24 16:16 model-00001-of-00006.safetensors
-rw-r----- 1 root root 4.6G Jun 24 16:16 model-00002-of-00006.safetensors
-rw-r----- 1 root root 4.6G Jun 24 16:16 model-00003-of-00006.safetensors
-rw-r----- 1 root root 4.7G Jun 24 16:16 model-00004-of-00006.safetensors
-rw-r----- 1 root root 4.6G Jun 24 16:17 model-00005-of-00006.safetensors
-rw-r----- 1 root root 3.0G Jun 24 16:17 model-00006-of-00006.safetensors
-rw------- 1 root root  32K Jun 24 16:16 modeling_baichuan.py
-rw-r----- 1 root root  23K Jun 24 16:17 model.safetensors.index.json
-rw------- 1 root root 9.0K Jun 24 16:16 quantizer.py
-rw-r----- 1 root root  544 Jun 24 16:17 special_tokens_map.json
-rw------- 1 root root 8.9K Jun 24 16:17 tokenization_baichuan.py
-rw-r----- 1 root root  918 Jun 24 16:17 tokenizer_config.json
-rw-r----- 1 root root 2.0M Jun 24 16:17 tokenizer.model

合并后的模型目录和原来的基础模型差不多，文件大小也差不多。

测试合并lora之后的模型文件

import torch
import torch_npu
from torch_npu.npu import amp
from torch_npu.contrib import transfer_to_npu
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers.generation.utils import GenerationConfig

model_path="/home/xxxx/baichuan-inc/Baichuan2-13B-Chat-lora-merge"  #这里用的是合并后的模型

tokenizer = AutoTokenizer.from_pretrained(model_path,
    revision="v2.0",
    use_fast=False,
    trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_path,
    revision="v2.0",
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True).half().npu().eval()
 
model.generation_config = GenerationConfig.from_pretrained(model_path, revision="v2.0")

messages = []
messages.append({"role": "user", "content": "讲一个100字左右的故事"})

response = model.chat(tokenizer, messages)
print(response)

其它说明

测试mindie过程发现模型保存为bfloat16会有问题，所以上述代码中我都是使用float16。

参考资料

LLM - LoRA 模型合并与保存