[Feature]: Model download can don't put to and read from HF cahe first?

[wikeeyang]

Feature description

项目是个好项目，兼容性也不错，作者们都挺用心，尽量让用户傻瓜型操作。
不过就是因为太傻瓜型，所以可能在模型下载、存储、或者手工配置这块，有点傻瓜，一定要在 HF Cache 缓存访问。。。
个人测试情况如下：
按照 Wiki 和 Issue 里的各种资料提示，手工将预先下载好的模型，放置到如下路径：
HiDream-Dev-NF4

和 Llama 3.1 8B Instruct 以及 NF4 模型

启动几次，设置好模型加载参数，“不预先加载” ， “不自动下载” ， Llama repo 改成：unsloth/Llama-3.1-8B-Instruct-bnb-4bit
N 次启动后，本地已经检测到有模型。且只有一个本地模型，如下：

加载模型跑图，后台依然要去 HF 从头开始下载。。。

为了测试这个 Cache 是否必须存在，试着完整下载了上述两个模型，启动跑图正常。。。
删除 Cache 的两个缓存目录，再跑图，又得重新下载。。。

这个模型管理机制是否能更改一下，先访问本地模型目录，没有的话，再去 HF 下载或补充下载缺失的部分？
否则，Cache 需要存一份，项目路径下的 models 目录下再管理保存一份，太傻瓜了。。。不管 SD WebUI 还是 ComfyUI ，都没这么弱智的设计啊。。。

Version Platform Description

Windows 11 x64 ，CUDA 12.4 ，Pytorch 2.60 ，Python 3.12

[wikeeyang]

[vladmandic]

English only please!

[wikeeyang]

English means: If I put model into project's models path manually at first, for example, put models--azaneko--HiDream-11-Dev-nf4 under "models\Diffusers" and put models--Meta--Llama-3.1-8B-Instruct and models--unsloth--Llama-3.1-8B-Instruct-bnb-4bit to "models\Text-encoder", and config the system model parameter to "not download auto", "not load auto", close all online model list and only one local model as above image disply, the click run, but the backend "modelloader.py" force to download the model and cache to HF cache directory again, after all downloaded model file to the cache path and refresh the models path, the text to image job can run normlly, but if I clean the HF cache file, restart the project, the backend will force download again, this configure is too simply, I think you can set the parameter to only download once a time is enough.....

[wikeeyang]

and you can translate whole page by google web translate easy, due to your project support Chinese, you can understand the image easy, the method of hf sanp_download script and all directory map is your project strandrd map.

[vladmandic]

for sdnext github issues english only is a rule, lets stay with that. anyhow.

re-download is automatically triggered by hf download method because it cannot determine version of manually downloaded model.
but yes, allowing to set local-files-only flag is a valid feature request and will be added.

not sure why tqdm-class and other settings are underlined in the screenshot?