---
title: Lora_2750steps_sanguo
canonical_url: "https://www.modelscope.cn/models/arronSs21/Lora_2750steps_sanguo"
md_url: "https://www.modelscope.cn/models/arronSs21/Lora_2750steps_sanguo.md"
repository: arronSs21/Lora_2750steps_sanguo
last_updated: 2026-01-20
base_model:
  - Tongyi-MAI/Z-Image-Turbo
base_model_relation: finetune
parameters: 85.0M
tensor_type:
  - BF16
library_name:
  - safetensors
supports_inference: txt2img
downloads: 8
stars: 0
---

# Lora_2750steps_sanguo

> Lora_2750steps_sanguo - arronSs21 在 ModelScope 开源的模型。三国

arronSs21/Lora_2750steps_sanguo 是 ModelScope 魔搭社区上的 85.0M 参数机器学习模型，基于 Tongyi-MAI/Z-Image-Turbo 构建，并支持在线推理（txt2img）。

- **Repository**: arronSs21/Lora_2750steps_sanguo
- **Parameters**: 85.0M
- **Base model**: Tongyi-MAI/Z-Image-Turbo
- **Online inference**: txt2img
- **Downloads**: 8
- **Stars**: 0
- **Last updated**: 2026-01-20

Source: https://www.modelscope.cn/models/arronSs21/Lora_2750steps_sanguo

---

# 第一步：安装必要依赖
!pip install modelscope peft transformers accelerate bitsandbytes -q

# 第二步：加载基础模型（魔搭版）
from modelscope import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel, LoraConfig

# 替换成魔搭上的基础模型 ID（如通义千问、Llama）
base_model_id = "qwen/Qwen-7B-Chat"
lora_path = "/home/model-user/work/your_lora.safetensors"  # 你的 LoRA 路径

# 加载基础模型和分词器
tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    trust_remote_code=True,
    device_map="auto",  # 自动分配显卡/CPU
    load_in_4bit=True,  # 节省显存，可选
)

# 第三步：加载 LoRA（核心步骤）
try:
    # 方式1：直接加载 LoRA 文件（推荐）
    model = PeftModel.from_pretrained(model, lora_path)
    print("LoRA 加载成功！")
except Exception as e:
    print("直接加载失败，尝试用配置加载：", e)
    # 方式2：手动配置 LoRA 后加载
    lora_config = LoraConfig.from_pretrained(lora_path)
    model = PeftModel.from_pretrained(model, lora_path, config=lora_config)

# 测试效果
inputs = tokenizer("你好", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))