train_lora.py

# train_lora.py – QLoRA + DeepSpeed DreamBooth Fine-Tuning (Stable Diffusion)

import os, argparse, torch
from diffusers import StableDiffusionPipeline, DDPMScheduler
from diffusers import DreamBoothLoraTrainer
from peft import LoraConfig
from accelerate import Accelerator

parser = argparse.ArgumentParser()
parser.add_argument("--data", default="./nyc_ads_dataset")  # 你的训练图片目录
args = parser.parse_args()

# LoRA 配置（兼容 QLoRA）
lora_cfg = LoraConfig(
    r=8,
    lora_alpha=32,
    lora_dropout=0.05,
    target_modules=["q_proj", "v_proj"]
)

# 4-bit 量化加载 SD-1.5
pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float16,
    load_in_4bit=True,
    quantization_config={
        "bnb_4bit_compute_dtype": torch.float16,
        "bnb_4bit_use_double_quant": True,
        "bnb_4bit_quant_type": "nf4"
    },
)

# DreamBooth LoRA Trainer
trainer = DreamBoothLoraTrainer(
    instance_data_root=args.data,
    instance_prompt="a photo of an urbanad nyc",
    lora_config=lora_cfg,
    output_dir="./nyc-ad-model",
    max_train_steps=400,
    train_batch_size=1,
    gradient_checkpointing=True,
)

# DeepSpeed ZeRO-3 加速 / 显存拆分
accelerator = Accelerator(
    mixed_precision="fp16",
    deepspeed_config="./ds_config_zero3.json"   # 需提前放置
)

# 开始训练
trainer.train(accelerator)