contrib: support modelscope community

[tastelikefeet]

Support gguf models of modelscope community.
linux/mac/windows HF/MS download tested.

[ngxson]

I think we need to re-think a bit about the implementation

From UX perspective, adding 4 more args may not be intuitive for most users, especially for most people outside china who won't use this (I understand the HF is not accessible from china checked with HF team - HF is still accessible in china, but still we should allow switching model host in a way that is less confused)

What I'm thinking is to handle the case where user can add protocol prefix to the existing -m, like what we have in llama-run. So for example

• hugging face: -m hf://user/model:quant (equivalent to -hf user/model:quant)
• modelscope: -m ms://user/model:quant

[ngxson]

What I'm thinking is to handle the case where user can add protocol prefix to the existing -m, like what we have in llama-run. So for example

* hugging face: `-m hf://user/model:quant` (equivalent to `-hf user/model:quant`)

* modelscope: `-m ms://user/model:quant`

Btw, we also have hf-mirror.com, so I think it's safe to say that adding one set of 4 arguments per host is not a scalable solution. So I would prefer to go with the protocol://... format

[foldl]

The abbreviation MS looks weird to me.

[tastelikefeet]

What I'm thinking is to handle the case where user can add protocol prefix to the existing -m, like what we have in llama-run. So for example

* hugging face: `-m hf://user/model:quant` (equivalent to `-hf user/model:quant`)

* modelscope: `-m ms://user/model:quant`

Btw, we also have hf-mirror.com, so I think it's safe to say that adding one set of 4 arguments per host is not a scalable solution. So I would prefer to go with the protocol://... format

This change was previously considered too significant for contributors(to add a new argument protocol for llama-cli and llama-server), so I chose a more conservative approach. Do we need to change to this way?

[ngxson]

This change was previously considered too significant for contributors(to add a new argument protocol for llama-cli and llama-server), so I chose a more conservative approach. Do we need to change to this way?

I still think having the protocol:// will be more extensible because:

• It's easier to add a third provider in the future
• We can also make it a positional arg (that's an idea for the future), so we can eliminate using -m and simply doing llama-cli protocol://user/model

Also, please correct if I'm wrong, but aren't modelscope and HF urls using the same structure? The format is: {base_domain}/resolve/main/{file}, then in this case I think a lot of codes in this PR is redundant, as we can simply have the base_domain replaceable.

[ngxson]

I would prefer a more simple approach. At the moment, this feature is kinda "nice-to-have", so we should start with something simple, then improve it later when more users use this.

[tastelikefeet]

This change was previously considered too significant for contributors(to add a new argument protocol for llama-cli and llama-server), so I chose a more conservative approach. Do we need to change to this way?

I still think having the protocol:// will be more extensible because:

• It's easier to add a third provider in the future
• We can also make it a positional arg (that's an idea for the future), so we can eliminate using -m and simply doing llama-cli protocol://user/model

Also, please correct if I'm wrong, but aren't modelscope and HF urls using the same structure? The format is: {base_domain}/resolve/main/{file}, then in this case I think a lot of codes in this PR is redundant, as we can simply have the base_domain replaceable.

No, the url may be a little different. HF is resolve/main, MS is resolve/master

[yingdachen]

I think a lot of discussions in this thread are quite constructive, which may result in a cleaner implementation that can facilitate llama.cpp users to access their needed model files from ModelScope. 

At the same time, there are obvious values to having another alternative community. For one thing, unlike lightweight open-source codes, large model weights require much more than a "mirror" site for efficient and high-speed access. ModelScope is dedicated to making models more accessible to our over 15-million users. Obviously integration into popular tools/frameworks like llama.cpp can greatly facilitate the efforts. We also believe easier access to wider model selection can benefit users of llama.cpp, and the open source communities as a whole. After all, the easy access to gguf models on huggingface via mode-id must have been implemented out of the same belief that such integration can connect llama.cpp better with the model-ecosystems. Otherwise, standalone operations like "wget" can get gguf models as well :) 

We will take the discussions today back to see how we can coordinate server-side protocol, to better accommodate client-side implementations that have been built around the hf ecosystem. @tastelikefeet

[tastelikefeet]

@ngxson @ggerganov We fixed the conversations and updated our backend code, and fixed some code that hard-coded Hugging Face endpoint to support ModelScope and other communities with the same downloading protocol, meanwhile we changed the HF_ENDPOINT to MODEL_ENDPOINT to make note developers with multiple communities support. cc @yingdachen

[ngxson]

Maybe better to move this function to arg.cpp, but I'll do it later

[ngxson]

Thanks for the contrib, I'll merge this once the CI is green