搭载MPS的MacOS运行Hugging Face Stable Diffusion模型报错求助

问题背景 #

在搭载MPS的macOS系统中运行基于Hugging Face的Stable Diffusion XL模型代码时，所有图形相关模型均触发相同维度不匹配错误，无法生成预期的PIL格式图像。

运行代码 #

import gradio as gr
import torch
import numpy as np
import modin.pandas as pd
from PIL import Image
from diffusers import DiffusionPipeline
import os

os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"

device = 'cuda' if torch.cuda.is_available() else 'mps'

if torch.cuda.is_available():
    PYTORCH_CUDA_ALLOC_CONF = {'max_split_size_mb': 8000}
    torch.cuda.max_memory_allocated(device=device)
    torch.cuda.empty_cache()

    pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16,
                                             variant="fp16", use_safetensors=True)
    pipe.enable_xformers_memory_efficient_attention()
    pipe = pipe.to(device)
    torch.cuda.empty_cache()

    refiner = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0", use_safetensors=True,
                                                torch_dtype=torch.float16, variant="fp16")
    refiner.enable_xformers_memory_efficient_attention()
    refiner = refiner.to(device)
    torch.cuda.empty_cache()

    upscaler = DiffusionPipeline.from_pretrained("stabilityai/sd-x2-latent-upscaler", torch_dtype=torch.float16,
                                                 use_safetensors=True)
    upscaler.enable_xformers_memory_efficient_attention()
    upscaler = upscaler.to(device)
    torch.cuda.empty_cache()
else:
    pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", use_safetensors=True)
    pipe = pipe.to(device)
    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
    refiner = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0", use_safetensors=True)
    refiner = refiner.to(device)
    refiner.unet = torch.compile(refiner.unet, mode="reduce-overhead", fullgraph=True)

n_steps = 40
high_noise_frac = 0.8

pipe.enable_attention_slicing()
PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0

def genie(prompt, negative_prompt, height, width, scale, steps, seed, upscaling, prompt_2, negative_prompt_2):
    generator = torch.Generator(device=device).manual_seed(seed)
    int_image = pipe(prompt, prompt_2=prompt_2, negative_prompt=negative_prompt, negative_prompt_2=negative_prompt_2,
                     num_inference_steps=steps, height=height, width=width, guidance_scale=scale,
                     num_images_per_prompt=1, generator=generator, output_type="latent").images
    if upscaling == 'Yes':
        image = \
            refiner(prompt=prompt, prompt_2=prompt_2, negative_prompt=negative_prompt,
                    negative_prompt_2=negative_prompt_2,
                    image=int_image).images[0]
        upscaled = upscaler(prompt=prompt, negative_prompt=negative_prompt, image=image, num_inference_steps=5,
                            guidance_scale=0).images[0]
        torch.cuda.empty_cache()
        return (image, upscaled)
    else:
        image = \
            refiner(prompt=prompt, prompt_2=prompt_2, negative_prompt=negative_prompt,
                    negative_prompt_2=negative_prompt_2,
                    image=int_image).images[0]
        torch.cuda.empty_cache()
    return (image, image)


gr.Interface(fn=genie, inputs=[gr.Textbox(
    label='What you want the AI to generate. 77 Token Limit. A Token is Any Word, Number, Symbol, or Punctuation. Everything Over 77 Will Be Truncated!'),
    gr.Textbox(label='What you Do Not want the AI to generate. 77 Token Limit'),
    gr.Slider(512, 1024, 768, step=128, label='Height'),
    gr.Slider(512, 1024, 768, step=128, label='Width'),
    gr.Slider(1, 15, 10, step=.25,
              label='Guidance Scale: How Closely the AI follows the Prompt'),
    gr.Slider(25, maximum=100, value=50, step=25, label='Number of Iterations'),
    gr.Slider(minimum=1, step=1, maximum=999999999999999999, randomize=True, label='Seed'),
    gr.Radio(['Yes', 'No'], value='No', label='Upscale?'),
    gr.Textbox(label='Embedded Prompt'),
    gr.Textbox(label='Embedded Negative Prompt')],
             outputs=['image', 'image'],
             title="Stable Diffusion XL 1.0 GPU",
             description="SDXL 1.0 GPU. <br><br><b>WARNING: Capable of producing NSFW (Softcore) images.</b>",
             article="If You Enjoyed this Demo and would like to Donate, you can send to any of these Wallets. <br>BTC: bc1qzdm9j73mj8ucwwtsjx4x4ylyfvr6kp7svzjn84 <br>3LWRoKYx6bCLnUrKEdnPo3FCSPQUSFDjFP <br>DOGE: DK6LRc4gfefdCTRk9xPD239N31jh9GjKez <br>SHIB (BEP20): 0xbE8f2f3B71DFEB84E5F7E3aae1909d60658aB891 <br>ETH: 0xbE8f2f3B71DFEB84E5F7E3aae1909d60658aB891 <br>Code Monkey: Manjushri").launch(
    debug=True, max_threads=80)

触发错误 #

step_index = (self.timesteps == timestep).nonzero().item()

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

RuntimeError: Expected dst.dim() >= src.dim() to be true, but got false. (Could this error message be improved? If so, please report an enhancement request to PyTorch.)

解决建议 #

• 禁用torch.compile：MPS后端对torch.compile的兼容性不足，这是引发维度错误的核心原因之一。修改MPS分支代码，移除对unet的编译逻辑：

  else:
      pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", use_safetensors=True)
      pipe = pipe.to(device)
      # 移除该行：pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
      refiner = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0", use_safetensors=True)
      refiner = refiner.to(device)
      # 移除该行：refiner.unet = torch.compile(refiner.unet, mode="reduce-overhead", fullgraph=True)
  

• 指定float32数据类型：MPS对float16的支持有限，强制使用float32可避免维度转换异常：

  else:
      pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", use_safetensors=True, torch_dtype=torch.float32)
      pipe = pipe.to(device)
      refiner = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-refiner-1.0", use_safetensors=True, torch_dtype=torch.float32)
      refiner = refiner.to(device)
  

• 统一timesteps设备：手动确保timesteps与模型在同一设备上，避免维度不匹配：
  在genie函数生成int_image前添加：

  pipe.timesteps = pipe.timesteps.to(device)
  refiner.timesteps = refiner.timesteps.to(device)
  

• 更新依赖库：确保torch、diffusers、transformers为最新兼容版本，修复已知MPS适配问题：

  pip install --upgrade torch diffusers transformers accelerate
  

• 关闭attention slicing（可选）：部分场景下enable_attention_slicing会与MPS后端冲突，尝试注释掉该行：

  # pipe.enable_attention_slicing()
  

内容的提问来源于stack exchange，提问作者B.AL