在 Mac 上本地部署文生图服务：Z-Image-Turbo 实践

pip install torch diffusers transformers accelerate gguf>=0.10.0

import torch
from diffusers import ZImagePipeline, ZImageTransformer2DModel, GGUFQuantizationConfig
from huggingface_hub import hf_hub_download

device = "mps" if torch.backends.mps.is_available() else "cpu"
dtype = torch.float16 if device == "mps" else torch.float32

# 下载量化模型
gguf_path = hf_hub_download(
    repo_id="jayn7/Z-Image-Turbo-GGUF",
    filename="z_image_turbo-Q4_K_S.gguf"
)

# 加载量化配置
quant_config = GGUFQuantizationConfig(compute_dtype=dtype)

transformer = ZImageTransformer2DModel.from_single_file(
    gguf_path,
    quantization_config=quant_config,
    torch_dtype=dtype,
)

# 构建 pipeline
pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    transformer=transformer,
    torch_dtype=dtype,
)
pipe.to(device)

image = pipe(
    prompt="一只可爱的柴犬，富士山背景，樱花飘落",
    height=512,  # Mac 建议 512，1024 可能内存不足
    width=512,
    num_inference_steps=4,  # 步数越少越快，4步约60秒
    guidance_scale=0.0,     # 必须设为 0
    generator=torch.Generator(device=device).manual_seed(42),
).images[0]

image.save("output.png")

# 简化的 MCP 服务框架
@mcp.tool()
def generate_image(
    prompt: str,
    width: int = 512,
    height: int = 512,
    steps: int = 4
) -> str:
    """Generate image from text prompt"""
    image = pipe(
        prompt=prompt,
        width=width,
        height=height,
        num_inference_steps=steps,
        guidance_scale=0.0,
    ).images[0]
    
    output_path = f"generated_{int(time.time())}.png"
    image.save(output_path)
    return output_path