gpt4 book ai didi

m4macmini本地部署ComfyUI,测试Flux-dev-GGUF的workflow模型10步出图,测试AI绘图性能,基于MPS(fp16),优点是能耗小和静音

转载 作者:撒哈拉 更新时间:2024-12-09 00:25:25 58 4
gpt4 key购买 nike

m4 mac mini已经发布了一段时间,针对这个产品,更多的是关于性价比的讨论,如果抛开各种补贴不论,价位上和以前发布的mini其实差别不大,真要论性价比,各种windows系统的mini主机的价格其实是吊打苹果的.

本次我们针对m4 mac mini的AI性能做个测试,使用目前泛用性最广的AI工作流软件:ComfyUI框架,基于MPS(fp16)模式进行测试.

Mac Os 本地部署ComfyUI

首先确保本机已经安装好了基于arm架构的Python3.11,之所以使用Python3.11,是因为这个版本性能有一定的优化,又不会像最新的3.13由于版本过新,引发依赖装不上的问题.

Mac版本Python3.11安装包的下载地址

https://python.org

随后克隆官方项目

git clone https://github.com/comfyanonymous/ComfyUI.git

接着安装 MPS 版本的 torch 。

pip3 install --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cpu

然后安装依赖:

pip3 install -r requirements.txt

依赖安装完毕后,需要升级一下SSL证书:

bash /Applications/Python*/Install\ Certificates.command

接着安装 ComfyUI 的 Manager 项目,用来安装各种节点:

cd custom_nodes  
git clone https://github.com/ltdrdata/ComfyUI-Manager.git

至此ComfyUI项目就部署好了.

Flux-dev-GGUF模型下载

下载需要的flux-dev模型,由于官方的模型体积太大(23G),这里我们下载GGUF的量化版本

https://pan.quark.cn/s/2907b57697fe

模型名称分别是:flux1-dev-Q4_1.gguf和t5-v1_1-xxl-encoder-Q5_K_M.gguf,将其分别放到models的UNET目录和clip目录.

随后,回到项目的根目录,输入命令,启动ComfyUI服务:

python3 main.py --force-fp16

这里强制使用fp16精度用来提升性能.

程序返回

liuyue@mini ComfyUI % python3 main.py --force-fp16  
[START] Security scan  
[DONE] Security scan  
## ComfyUI-Manager: installing dependencies done.  
** ComfyUI startup time: 2024-12-08 23:04:08.464703  
** Platform: Darwin  
** Python version: 3.11.9 (v3.11.9:de54cf5be3, Apr  2 2024, 07:12:50) [Clang 13.0.0 (clang-1300.0.29.30)]  
** Python executable: /Library/Frameworks/Python.framework/Versions/3.11/bin/python3  
** ComfyUI Path: /Volumes/ssd/work/ComfyUI  
** Log path: /Volumes/ssd/work/ComfyUI/comfyui.log  
  
Prestartup times for custom nodes:  
   0.7 seconds: /Volumes/ssd/work/ComfyUI/custom_nodes/ComfyUI-Manager  
  
Total VRAM 24576 MB, total RAM 24576 MB  
pytorch version: 2.5.1  
Forcing FP16.  
Set vram state to: SHARED  
Device: mps  
Using sub quadratic optimization for cross attention, if you have memory or speed issues try using: --use-split-cross-attention  
[Prompt Server] web root: /Volumes/ssd/work/ComfyUI/web  
### Loading: ComfyUI-Manager (V2.51.9)  
### ComfyUI Revision: 2859 [b4526d3f] | Released on '2024-11-24'  
[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/alter-list.json  
[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/model-list.json  
[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/github-stats.json  
[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/custom-node-list.json  
Torch version 2.5.1 has not been tested with coremltools. You may run into unexpected errors. Torch 2.4.0 is the most recent version that has been tested.  
[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/extension-node-map.json  
  
Import times for custom nodes:  
   0.0 seconds: /Volumes/ssd/work/ComfyUI/custom_nodes/websocket_image_save.py  
   0.0 seconds: /Volumes/ssd/work/ComfyUI/custom_nodes/ComfyUI-GGUF  
   0.1 seconds: /Volumes/ssd/work/ComfyUI/custom_nodes/ComfyUI-Manager  
   2.2 seconds: /Volumes/ssd/work/ComfyUI/custom_nodes/ComfyUI-MLX  
  
Starting server  
  
To see the GUI go to: http://127.0.0.1:8188

代表部署成功,访问:http://127.0.0.1:8188 。

测试Flux-dev-GGUF工作流

下载基于GGUF的工作流:

https://promptingpixels.com/flux-gguf/

导入工作流后,输入提示词

a super sexy gal holding a sign that says "ComfyUI Mac"

意思是性感女子举着一个牌子,上面写着 ComfyUI Mac 。

此时,可以直接执行工作流,程序返回:

ggml_sd_loader:  
 13                            144  
 0                              50  
 14                             25  
Requested to load FluxClipModel_  
Loading 1 new model  
loaded completely 0.0 323.94775390625 True  
Requested to load FluxClipModel_  
Loading 1 new model  
  
ggml_sd_loader:  
 1                             476  
 3                             304  
model weight dtype torch.bfloat16, manual cast: None  
model_type FLUX  
Requested to load Flux  
Loading 1 new model  
loaded completely 0.0 7181.8848876953125 True  
 20%|██████████████████▌                                                                          | 2/10 [01:04<04:18, 32.27s/it]

每秒的迭代稳定在30次左右,一张图大概需要3-5分钟左右.

笔者的 m4 mac mini 的配置是丐版升级到24G内存,在出图的过程中,通过活动监视器可知,内存没有被占满:

可以看到,只使用了21G的内存,有网友使用纯丐版16G内存的mini进行测试,16g内存实际刨除系统占用,空闲最多也就10g,超出的部分只能跑SSD的虚拟内存,导致GPU跑不满,所以丐版16G内存是有可能导致出图的效率降低.

最后是10步迭代的出图效果: