Views
No views yet
1{
2 "peft_type": "LORA",
3 "base_model_name_or_path": "/home/lbn/model/Peach-2.0-9B-8k-Roleplay",
4 "r": 32,
5 "lora_alpha": 64,
6 "lora_dropout": 0.1,
7 "target_modules": [
8 "q_proj",
9 "k_proj",
10 "v_proj",
11 "o_proj",
12 "gate_proj",
13 "down_proj",
14 "up_proj"
15 ],
16 "bias": "none",
17 "task_type": "CAUSAL_LM"
18}1from transformers import AutoModelForCausalLM, AutoTokenizer
2from peft import PeftModel
3import torch
4
5# 加载基础模型和分词器
6base_model_path = "LBN154/Peach-2.0-9B-8k-Roleplay"
7tokenizer = AutoTokenizer.from_pretrained(base_model_path)
8model = AutoModelForCausalLM.from_pretrained(
9 base_model_path,
10 torch_dtype=torch.float16,
11 device_map="auto"
12)
13
14# 加载 LoRA 适配器
15lora_path = "LBN154/Peach-2.0-9B-8k-Roleplay-urban-romance-lora" # 替换为你的 HF 仓库名
16model = PeftModel.from_pretrained(model, lora_path)
17
18# 对话示例
19messages = [
20 {"role": "system", "content": "李清，26岁的都市白领，刚失恋。"},
21 {"role": "user", "content": "好久不见！李清~"}
22]
23
24input_ids = tokenizer.apply_chat_template(
25 conversation=messages,
26 tokenize=True,
27 return_tensors="pt",
28 add_generation_prompt=True
29).to(model.device)
30
31output = model.generate(
32 input_ids=input_ids,
33 max_new_tokens=512,
34 temperature=0.5,
35 top_p=0.7,
36 repetition_penalty=1.05
37)
38
39response = tokenizer.decode(output[0], skip_special_tokens=True)
40print(response)lora/
├── adapter_config.json # LoRA 配置
├── adapter_model.safetensors # LoRA 权重
├── tokenizer.json # 分词器配置
├── tokenizer_config.json # 分词器配置
├── special_tokens_map.json # 特殊 token 映射
└── training_args.bin # 训练参数transformers >= 4.37.0peft >= 0.18.0torch >= 2.0.0accelerate (可选，用于多 GPU)