[Feature]: Modelscope alternative download

[alerikaisattera]

Feature description

Add downloading models from Modelscope as an alternative to Huggingface

Advantages:

• MS can be used instead of HF when connection to HF CDN fails for various reasons
• Some models that are gated on HF are public on MS

Disadvantages:

• Some models do not have repos at MS
• SD.Next seems to be highly integrated with HF and adding alternatives is very difficult

Version Platform Description

No response

[vladmandic]

valid, will add in the future

[vladmandic]

modelscope cannot be used instead of huggingface as-is, it requires implementation of modelscope api.
however, current state of modelscope api is that it lacks even basic search capabilities - it's simply not implemented yet.
you can see current api implementation at https://github.com/modelscope/modelscope/blob/master/modelscope/hub/api.py
its pretty much trimmed down and customized version of huggingface api. unfortunately, they trimmed down some highly required features making it unusable.

i'll keep this feature request open for now, but i'm not going to write massive code to add modelscope downloader if they don't provide basic api functionality.

[alerikaisattera]

Could it be possible to impement just what's possible now? Alternative model download links (for use when HF CDN connection fails) don't require search

[vladmandic]

download from hf is done using hf api not using raw links.
so cant just replace links as hf api cannot work with modelscope.
which means i'd have to implement modelscope api - which i'm wiling to do if their api has minimum required functionality.

[alerikaisattera]

MS API seems to have download

[vladmandic]

yes, it does. it does not have search. nothing to be gained here by repeating it.
my workflow requires search as first step and I'm NOT going to reinvent the wheel just for model scope.

if you want to persuade someone, talk to them and ask them why list_models api is so limited?

[alerikaisattera]

Not that I know anything about either HF or MS API to tell them about it. For now this may be put into backlog. Assuming that models downloaded from MS manually and placed into respective SD.Next directories do work, this can wait

[vladmandic]

they should. and ideally, you should be using modelscope cli for that, not manual download using browser.
for example, something like this should work:

pip install modelscope
modelscope download --model Tongyi-MAI/Z-Image-Turbo --local_dir /sdnext/models/Diffusers