import os

os.environ["CUDA_VISIBLE_DEVICES"] = ""
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch


# 由于 LoRA 权重不能直接转 GGUF，你必须先将它与原模型合并。

lora_path = "model/qwen-0.8b-tender-lora/checkpoint-430"
save_path = "model/qwen-merged-model"

# 请将此处替换为你上面 find 命令查出来的真实绝对路径
base_model_path = "/home/user/.cache/huggingface/hub/models--Qwen--Qwen3.5-0.8B/snapshots/2fc06364715b967f1860aea9cf38778875588b17"

print(f"正在尝试从本地路径加载模型: {base_model_path}")

# 显式指定从本地加载
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_path,
    torch_dtype=torch.float32,
    device_map={"": "cpu"},
    local_files_only=True, # 强制不联网，路径不对会立刻报错
    trust_remote_code=True
)

model = PeftModel.from_pretrained(base_model, lora_path)
merged_model = model.merge_and_unload()

merged_model.save_pretrained(save_path)
tokenizer = AutoTokenizer.from_pretrained(base_model_path)
tokenizer.save_pretrained(save_path)
print("模型已成功合并并保存。")