#!/bin/bash
# 上传脚本 - 合并 adapter 并上传到 HuggingFace Hub
# 适用于 Qwen3.5-0.8B-HOS 训练流程

set -e

echo "============================================"
echo "  上传模型到 HuggingFace Hub"
echo "  目标: lxcxjxhx/Qwen3.5-0.8B-HOS"
echo "============================================"

# 配置
OUTPUT_DIR="outputs/qwen35-0.8b-cybersec-qlora"
MERGED_DIR="outputs/qwen35-0.8b-merged"
HUB_MODEL_ID="lxcxjxhx/Qwen3.5-0.8B-HOS"

# 检查训练输出
if [ ! -d "$OUTPUT_DIR" ]; then
    echo "错误: 训练输出目录不存在: $OUTPUT_DIR"
    exit 1
fi

# Step 1: 合并 adapter
echo "[1/3] 合并 LoRA adapter..."
python -c "
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

print('加载基础模型...')
base_model = AutoModelForCausalLM.from_pretrained(
    'lxcxjxhx/Qwen3.5-0.8B-HOS',
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)

print('加载 LoRA adapter...')
model = PeftModel.from_pretrained(base_model, '$OUTPUT_DIR')

print('合并模型...')
model = model.merge_and_unload()

print('保存合并后的模型...')
model.save_pretrained('$MERGED_DIR')

print('保存 tokenizer...')
tokenizer = AutoTokenizer.from_pretrained('lxcxjxhx/Qwen3.5-0.8B-HOS', trust_remote_code=True)
tokenizer.save_pretrained('$MERGED_DIR')

print('合并完成!')
"

# Step 2: 推送到 HuggingFace Hub
echo "[2/3] 推送到 HuggingFace Hub..."
python -c "
from huggingface_hub import HfApi

api = HfApi()

print('上传模型到: $HUB_MODEL_ID')
api.upload_folder(
    folder_path='$MERGED_DIR',
    repo_id='$HUB_MODEL_ID',
    repo_type='model',
    commit_message='Upload Qwen3.5-0.8B-HOS merged model'
)

print('上传完成!')
"

# Step 3: 验证上传
echo "[3/3] 验证上传..."
python -c "
from huggingface_hub import model_info

info = model_info('$HUB_MODEL_ID')
print(f'模型: {info.modelId}')
print(f'最后更新: {info.lastModified}')
print(f'标签: {info.tags}')
print('验证完成!')
"

echo "============================================"
echo "  上传完成!"
echo "  模型地址: https://huggingface.co/$HUB_MODEL_ID"
echo "============================================"