如何在Hugging Face Transformers中合并微调Adapter与预训练模型并推送至Hub

目标1：合并预训练模型与Adapter并推送至Hugging Face Hub #

要将预训练模型与Adapter权重合并为完整模型文件并推送，核心是通过merge_and_unload()方法把Adapter权重整合到主模型中，具体操作如下：

1. 加载完整精度的预训练模型与Adapter
   注意：合并操作不能用8bit加载模型，必须加载完整精度版本：

   import torch
   from transformers import LlamaForCausalLM, LlamaTokenizer
   
   model_id = "./models_hf/7B"
   adapter_id = "./tmp/llama-output"
   
   tokenizer = LlamaTokenizer.from_pretrained(model_id)
   model = LlamaForCausalLM.from_pretrained(
       model_id,
       load_in_8bit=False,
       device_map='auto',
       torch_dtype=torch.float16
   )
   
   # 加载Adapter权重
   model.load_adapter(adapter_id)
   

2. 合并权重并保存/推送
   调用merge_and_unload()完成权重整合，之后可选择本地备份或推送至Hub：

   # 合并Adapter与主模型权重
   model = model.merge_and_unload()
   
   # 可选：保存合并后的模型到本地
   model.save_pretrained("./merged-llama-2-7B")
   tokenizer.save_pretrained("./merged-llama-2-7B")
   
   # 推送至Hugging Face Hub
   model.push_to_hub("myrepo/llama-2-7B-ft-summarization")
   tokenizer.push_to_hub("myrepo/llama-2-7B-ft-summarization")
   

   合并后的模型结构与原始预训练模型一致，推送后可像普通大模型直接加载使用。

目标2：分别推送预训练模型与Adapter，并组合加载推理 #

步骤1：分别推送至Hub #

推送预训练模型

直接将本地预训练模型推送到你的Hub仓库：

from transformers import LlamaForCausalLM, LlamaTokenizer

model_id = "./models_hf/7B"
hub_base_repo = "myrepo/llama-2-7B-base"

tokenizer = LlamaTokenizer.from_pretrained(model_id)
model = LlamaForCausalLM.from_pretrained(model_id)

tokenizer.push_to_hub(hub_base_repo)
model.push_to_hub(hub_base_repo)

推送Adapter权重

借助PEFT库将Adapter推送到独立的Hub仓库：

from peft import PeftModel
from transformers import LlamaForCausalLM

model_id = "./models_hf/7B"
adapter_id = "./tmp/llama-output"
hub_adapter_repo = "myrepo/llama-2-7B-summarization-adapter"

# 加载预训练模型与Adapter
base_model = LlamaForCausalLM.from_pretrained(model_id)
peft_model = PeftModel.from_pretrained(base_model, adapter_id)

# 推送Adapter到Hub
peft_model.push_to_hub(hub_adapter_repo)

步骤2：从Hub组合加载并推理 #

用Hub仓库地址替换本地路径，即可实现和本地一致的组合加载：

import torch
from transformers import LlamaForCausalLM, LlamaTokenizer
from peft import PeftModel

# Hub仓库地址
model_id = "myrepo/llama-2-7B-base"
adapter_id = "myrepo/llama-2-7B-summarization-adapter"

# 加载预训练模型与tokenizer
tokenizer = LlamaTokenizer.from_pretrained(model_id)
model = LlamaForCausalLM.from_pretrained(
    model_id,
    load_in_8bit=True,
    device_map='auto',
    torch_dtype=torch.float16
)

# 加载Adapter
model = PeftModel.from_pretrained(model, adapter_id)

# 推理示例
inputs = tokenizer("请输入你的测试文本", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=150)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

内容的提问来源于stack exchange，提问作者Aun Zaidi