ssh -p <你的端口> root@region-xx.autodl.com

# 1. 激活平台自带的 conda 环境
source /root/miniconda3/etc/profile.d/conda.sh && conda activate base

# 2. 开启平台学术加速，并安装支持 Qwen-Image-2.1 的 diffusers 主线 + 推理依赖
source /etc/network_turbo
pip install -q git+https://github.com/huggingface/diffusers.git \
  transformers accelerate sentencepiece protobuf torchvision gradio modelscope

python -c "import torch, diffusers; print(torch.__version__, torch.cuda.get_device_name(0), diffusers.__version__)"

OSError: Not enough disk space / No space left on device

export MODELSCOPE_CACHE=/root/autodl-tmp/modelscope_cache
export HF_HOME=/root/autodl-tmp/hf_cache
mkdir -p /root/autodl-tmp/modelscope_cache

cat >> /root/.bashrc <<'EOF'
export MODELSCOPE_CACHE=/root/autodl-tmp/modelscope_cache
export HF_HOME=/root/autodl-tmp/hf_cache
EOF

import os
import torch

# 铁律：强制把缓存落在 50G 数据盘，杜绝系统盘写满
os.environ["MODELSCOPE_CACHE"] = "/root/autodl-tmp/modelscope_cache"
os.environ["HF_HOME"] = "/root/autodl-tmp/hf_cache"

from modelscope import snapshot_download
from diffusers import QwenImage21Pipeline

# 1. 走国内高速通道缓存权重（免梯子，实测 50~100MB/s）
print("正在通过 ModelScope 通道拉取 Qwen-Image-2.1 完整权重...")
model_dir = snapshot_download("Qwen/Qwen-Image-2.1",
                             cache_dir="/root/autodl-tmp/modelscope_cache")

# 2. 载入满血 BF16 管道（显存占用约 28~31GB，32GB 卡可承受）
pipe = QwenImage21Pipeline.from_pretrained(model_dir, torch_dtype=torch.bfloat16).to("cuda")

# 3. 出图：30 步 + true_cfg 4.0，直接拿 RGBA 透明底
prompt = ("This is an RGBA image with transparency. A high-fashion futuristic "
          "cyberpunk editorial portrait, translucent holographic trench coat, "
          "glowing neon cyan circuit etchings, wet glass reflections, clean "
          "transparent background, 8k, award-winning aesthetics")

image = pipe(prompt=prompt,
             num_inference_steps=30,
             true_cfg_scale=4.0,
             width=1024, height=1024).images[0]

image.save("qwen_rgba.png")           # PNG 保留 alpha 通道
print(f"已保存 qwen_rgba.png，模式={image.mode}，尺寸={image.size}")

cd /root/autodl-tmp && python gen.py

import os
import torch
import gradio as gr

os.environ["MODELSCOPE_CACHE"] = "/root/autodl-tmp/modelscope_cache"
os.environ["HF_HOME"] = "/root/autodl-tmp/hf_cache"

from modelscope import snapshot_download
from diffusers import QwenImage21Pipeline

model_dir = snapshot_download("Qwen/Qwen-Image-2.1",
                             cache_dir="/root/autodl-tmp/modelscope_cache")
pipe = QwenImage21Pipeline.from_pretrained(model_dir, torch_dtype=torch.bfloat16).to("cuda")

def generate(prompt, steps, cfg, width, height):
    return pipe(prompt=prompt,
                num_inference_steps=int(steps),
                true_cfg_scale=float(cfg),
                width=int(width), height=int(height)).images[0]

with gr.Blocks(title="阿里 Qwen-Image-2.1 极速出图大厅 (RTX 4080 SUPER 32GB)") as demo:
    gr.Markdown("# 阿里 Qwen-Image-2.1 极速出图大厅\n"
                "**实机环境：AutoDL 西北B区 · vGPU-32GB · Diffusers 原生推理**")
    with gr.Row():
        with gr.Column():
            prompt_input = gr.Textbox(
                label="Prompt（支持原生 RGBA 透明通道）",
                value="This is an RGBA image with transparency. A high-fashion futuristic "
                      "cyberpunk editorial portrait, clean transparent background, 8k",
                lines=4)
            steps_slider = gr.Slider(10, 50, value=30, step=1, label="去噪步数 (Inference Steps)")
            cfg_slider = gr.Slider(1.0, 10.0, value=4.0, step=0.5, label="引导强度 (True CFG Scale)")
            width_slider = gr.Slider(512, 2048, value=1024, step=64, label="宽")
            height_slider = gr.Slider(512, 2048, value=1024, step=64, label="高")
            btn = gr.Button("立即出图 (Generate)", variant="primary")
        with gr.Column():
            output_img = gr.Image(label="生成结果预览")
    btn.click(generate,
              inputs=[prompt_input, steps_slider, cfg_slider, width_slider, height_slider],
              outputs=output_img)

# 只监听本机回环，靠 SSH 隧道访问，不要把端口裸奔到公网
demo.launch(server_name="127.0.0.1", server_port=6006)

python /root/autodl-tmp/app.py

ssh -CNg -L 6006:127.0.0.1:6006 root@<你的SSH域名> -p <你的SSH端口>

pipe = QwenImage21Pipeline.from_pretrained("unsloth/Qwen-Image-2.1-FP8",
                                          torch_dtype=torch.bfloat16).to("cuda")

AutoDL 西北B区 + vGPU-32GB + PyTorch2.5/CUDA12.4 镜像
  → source /etc/network_turbo && pip install git+diffusers
  → export HF_HOME / MODELSCOPE_CACHE 到 /root/autodl-tmp
  → modelscope 拉 Qwen/Qwen-Image-2.1
  → QwenImage21Pipeline(bf16) 出 RGBA 图 / 或 Gradio WebUI
  → 玩完立刻关机，传图用无卡开机