[GH-ISSUE #9610] Documents on how to convert a selftrained llava model into ollama supported gguf model file #52781

Open
opened 2026-04-29 00:51:29 -05:00 by GiteaMirror · 3 comments
Owner

Originally created by @lucasjinreal on GitHub (Mar 9, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/9610

Documents on how to convert a selftrained llava model into ollama supported gguf model file

Originally created by @lucasjinreal on GitHub (Mar 9, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/9610 Documents on how to convert a selftrained llava model into ollama supported gguf model file
GiteaMirror added the feature request label 2026-04-29 00:51:29 -05:00
Author
Owner

@pdevine commented on GitHub (Mar 12, 2025):

Unfortunately llava is a bit tricky because of the projector and the text model being split by default in the old llama.cpp engine. The new ollama engine will make this a bit easier, but we don't yet have support for llava (only mllama as of writing this comment, although more multimodal architectures are coming imminently!).

<!-- gh-comment-id:2716146636 --> @pdevine commented on GitHub (Mar 12, 2025): Unfortunately llava is a bit tricky because of the projector and the text model being split by default in the old llama.cpp engine. The new ollama engine will make this a bit easier, but we don't yet have support for llava (only mllama as of writing this comment, although more multimodal architectures are coming imminently!).
Author
Owner

@lucasjinreal commented on GitHub (Mar 12, 2025):

thank u, may I further ask what's the difference of use ollama in terms of llava compare with llama.cpp.
I suppose ollama is a golang wrapper of llama.cpp.

It ollama is easier, please kindly written some tutorials on how to convert any customized llava to support (such as different vit, difference projector design etc), I read llama.cpp's conversion, it's way toooo complicated and not really necessary IMO.

<!-- gh-comment-id:2716224793 --> @lucasjinreal commented on GitHub (Mar 12, 2025): thank u, may I further ask what's the difference of use ollama in terms of llava compare with llama.cpp. I suppose ollama is a golang wrapper of llama.cpp. It ollama is easier, please kindly written some tutorials on how to convert any customized llava to support (such as different vit, difference projector design etc), I read llama.cpp's conversion, it's way toooo complicated and not really necessary IMO.
Author
Owner

@pdevine commented on GitHub (Mar 12, 2025):

As I mentioned, llava works with the old llama.cpp engine through clip.cpp. New multimodal models will work in the new ollama engine (i.e. it does not wrap llama.cpp and the implementation is done in Ollama itself). You can find the model definitions in model/models/*.

<!-- gh-comment-id:2716544565 --> @pdevine commented on GitHub (Mar 12, 2025): As I mentioned, llava works with the old llama.cpp engine through clip.cpp. New multimodal models will work in the new ollama engine (i.e. it does *not* wrap llama.cpp and the implementation is done in Ollama itself). You can find the model definitions in `model/models/*`.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#52781