Added magic for file types

[jboero]

Add this to your /etc/magic file to enable the file command.

jboero@xps ~/Downloads> file *llama*
codellama-7b.Q8_0.gguf:     GGUF LLM model version=1
llama-2-7b.ggmlv3.q8_0.bin: GGML/GGJT LLM model version=3
llama-cpp.srpm.spec:        ASCII text

[KerfuffleV2]

I don't think this would go in grammers/, that directory is for the grammar sampling stuff.

file is a Unix tool, you probably shouldn't just add a config file for it with no explanation.

Also, your file configuration here doesn't actually have enough information to answer my question in #2990 about what type of GGML file you actually have. Knowing it's some kind of GGML/GGJT/GGMF file of a version isn't specific enough.

Also, at least on Linux file supports -m which lets you set a config file with info for these types. No need to added it to the system wide /etc/magic

---

I think if you wanted to add this it would make more sense to create a document in docs/ or something where you can show the example and include some information so people know how to use it and why they'd want to.

[jboero]

Okay I can do that until it gets into upstream magic. I'm also working in a branch with a dedicated tool to give more details about the file contents but this was low hanging fruit for now.

[jboero]

İs there an official mime type for models?

[staviq]

İs there an official mime type for models?

Funnily enough, there is an official "model" mime category, so a proper mime type (apparently it's deprecated name and now that are called media types, instead of mime types) would be

model/ggml
model/gguf

Registering new types "officially" seems to be done through an individual review by IANA.

This article links registration procedures near the very top of the page: https://www.iana.org/assignments/media-types/media-types.xhtml#model

From what I can tell, a media type, can include identifying magic.

Edit: So apparently, the "Linux" way of adding your own mime type is through XML files, added to one of $XDG_DATA_DIRS and followed by update-mime-database. This article explains it a bit better, it's like 10 years old, but it does appear at least Arch still does it this way: https://www.freedesktop.org/wiki/Specifications/AddingMIMETutor/

And I think considering the above, this should probably be moved to *.srpm.spec and/or other places directly involved in packaging llamacpp.

I don't think direct edits of Linux mime database are supported anywhere, as update-mime-database appears to be overriding it.

[jboero]

Yes I had planned to add to the RPM specs but will move it to docs for now and I need to craft it carefully enough to support the latest Fedora and the oldest *EL. Testers appreciated.

Funny mixing LLM models with 3D models in the model/* category.