Feat/add modelscope -ms args support

[yrk111222]

Overview

This PR adds ModelScope (-ms) option support. Based on the feedback #22279 about keeping responsibilities separated and following the hf-cache.cpp pattern

Additional information

Key Design Decisions

1. Downloading, quantization selection, and filtering logic stays in download.cpp

Following the reviewer's request, we did not duplicate these functions in ms-cache.cpp. Instead, the existing functions in download.cpp were templatized so they can be shared by both HF and MS. This keeps ms-cache.cpp focused on caching only — mirroring hf-cache.cpp's responsibilities. The templates are defined entirely within download.cpp, with no header changes, so HF behavior is unaffected.

2. ms-cache.cpp follows the pattern of hf-cache.cpp closely

ms_file struct mirrors hf_file (identical fields)
Public API matches hf-cache.h (get_repo_files, get_cached_files, finalize_file)
Same blob/snapshot/symlink cache structure
The only differences are platform-specific (unavoidable):
Auth: Cookie vs Bearer
Endpoint: dedicated ModelScope endpoint
Snapshot key: "master" (MS API has no single repo-wide commit hash)
oid: Sha256 from API

3. Cache path aligned with HF + token-less loading

MS uses refs/master as the snapshot key. Since MS files use regular filenames requiring ETag validation (unlike HF's content-addressed blobs), cached MS models need extra handling to load without a token — just like HF. When all files are cached, download tasks are skipped and cached paths are returned directly.

Usage

# Auto-select best model
llama-cli -ms Qwen/Qwen3-0.6B-GGUF -p "hello"

# Private repo
export MS_TOKEN=<token>
llama-cli -ms <owner>/<private-model>

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES

I used an LLM to translate the PR description and do some code review.

[ggml-gh-bot]

Hi @yrk111222, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[yrk111222]

HI @angt:
Rebased onto the latest master and resolved all conflicts. PR is now ready for review — I'd really appreciate your feedback or any suggestions!😊

[angt]

@yrk111222 That looks much better indeed, well done!
I'll try to do the the final review this week 👍

[yrk111222]

Hi @angt,

I noticed the recent large refactor to the download code, and this PR has fallen a bit behind. To minimize rework, would you rather I rebase it against the current state now, or hold off until things stabilize? Happy to follow your lead on this—just let me know what works best.

Thanks!😊

[angt]

Hi @angt,

I noticed the recent large refactor to the download code, and this PR has fallen a bit behind. To minimize rework, would you rather I rebase it against the current state now, or hold off until things stabilize? Happy to follow your lead on this—just let me know what works best.

Thanks!😊

Yes we are sorry for that, and I think we still need some refactoring.
But i believe that rebasing at each step might be easier to follow the process for you 🤔

[yrk111222]

Hi @angt,

Rebased onto the latest master and synced the MS code with the download refactor. Build + tests pass.
Whenever you have time, could you take another look and let me know if there's anything that still needs changing?

Thanks!😊

[yrk111222]

Hi @angt, rebased on latest master and aligned MS with the HF download pipeline changes. Key updates since last review:

• Aligned dflash/eagle3 sidecar support for MS
• Fixed server.cpp to include ms_repo in models_handler init (was HF-only)
• Fixed app/download.cpp to include ms_repo in has_source check
• Added ms_cache::remove_cached_repo and extended cache list/remove to MS

Would appreciate another review when you have a moment.😊

[yingdachen]

@angt could you please take another look and (hopefully) get this merged soon? this has been outstanding for a (unfortunate) long period of time...

thanks!

[yingdachen]

a gentle ping again @angt

hope we can get this merged before we need to resolve conflict again ... :(

[yrk111222]

Hi @angt, gentle ping again. Rebased on latest master and resolved conflicts. Is the refactoring stabilized enough for another PR review? Happy to address any feedback promptly... : )

[yingdachen]

this change has been pending for over two months (if we count the original implementation #22279 , then it is over 5 months), with multiple merges to keep it up to date.

is there any outstanding issue that is blocking the merge? we would really love to see things moving forward if the code is now in good shape :)

thanks for your time. @angt

[yrk111222]

Hi @angt,
sorry to bother you. Just synced ms-cache.cpp with the latest hf-cache.cpp refactoring (fs::path, fs_path_to_utf8, fs_write_atomic, etc.). This PR's feature is also widely requested by users: modelscope/modelscope#1658.
Would appreciate a review when you have time.😊