Loading with ggml llava through llama.cpp (PR)

#1
by cmp-nct - opened
This comment has been hidden
cmp-nct changed discussion title from Can you provide the full CLIP model, not just the extracted vision part ? to never mind
cmp-nct changed discussion status to closed

@cmp-nct can you let me know how you fixed this issue ?
image.png

cmp-nct changed discussion title from never mind to Loading with ggml llava

@cmp-nct can you let me know how you fixed this issue ?
image.png

Sure, I am using ggml llava inference which loads the CLIP model first for conversion, it expects a full model to extract the vision from it but this is already the vision only part.

I wrote a patch: https://github.com/ggerganov/llama.cpp/pull/4172

Anything that looks for "clip_vision_model" is code that expects a full CLIP, in that case you just need to skip the extraction code.

cmp-nct changed discussion status to open
cmp-nct changed discussion title from Loading with ggml llava to Loading with ggml llava through llama.cpp (PR)

Sign up or log in to comment