Support for multi-modal models

[rlancemartin]

I see LLama.cpp is working on multi-modal models like LLaVA:
ggerganov/llama.cpp#3436

Model is here:

2ab9be51b7dc737136b38093316a4d3577d1fb96281f1589adac7841f5b81c43  ../models/ggml-model-q5_k.gguf
b7c8ff0f58fca47d28ba92c4443adf8653f3349282cb8d9e6911f22d9b3814fe  ../models/mmproj-model-f16.gguf

Testing:

$ mkdir build && cd build && cmake ..
$ cmake --build .
$ ./bin/llava -m ../models/ggml-model-q5_k.gguf --mmproj ../models/mmproj-model-f16.gguf --image ~/Desktop/Papers/figure-3-1.jpg

Appears to add some new params:

--mmproj MMPROJ_FILE  path to a multimodal projector file for LLaVA. see examples/llava/README.md
--image IMAGE_FILE    path to an image file. use with multimodal models

It would be awesome if we can support in llama-cpp-python.

[Josh-XT]

That PR has been merged as of about an hour ago now.

[rlancemartin]

@abetlen any thoughts on this?

[damian0815]

i'm working on a strategy for supporting this by adding base64 image support to llama.cpp - this will mean the image can just be passed within the prompt. cf https://github.com/oobabooga/text-generation-webui/tree/main/extensions/multimodal

[marscod]

I have the latest version 0.2.11, with built llava from lama.cpp and getting this error:

from llama_cpp import (Llama, clip_model_load, llava_image_embed_make_with_filename, llava_image_embed_make_with_bytes,
ImportError: cannot import name 'clip_model_load' from 'llama_cpp'

[sagar-kris]

@marscod v0.2.11 doesn't have the latest merged changes yet, try

pip uninstall llama-cpp-python -y

and then to install with the latest merged changes

pip install -U --no-cache-dir llama-cpp-python@git+https://github.com/abetlen/llama-cpp-python.git@ef03d77b59718f7d422f24b15c653c2a57b087f3

[Josh-XT]

@marscod v0.2.11 doesn't have the latest merged changes yet, try

pip uninstall llama-cpp-python -y

and then to install with the latest merged changes

pip install -U --no-cache-dir llama-cpp-python@git+https://github.com/abetlen/llama-cpp-python.git@ef03d77b59718f7d422f24b15c653c2a57b087f3

It hasn't been merged yet. #821

[z3ugma]

I followed the direction to install at that commit ef03d77b5, how do I specify to use the llava instance and pass the mmproj file as an argument when using in Python code

from llama_cpp import Llama
llm = Llama(model_path="/Users/fred/llama.cpp/models/ggml-model-q4_k.gguf.1", mmproj=)

   from llama_cpp import (
ImportError: cannot import name 'clip_model_load' from 'llama_cpp' (venv/lib/python3.11/site-packages/llama_cpp/__init__.py)

[rlancemartin]

I talked to @ggerganov: what is missing is a Python API to run the CLIP model. There are plans for CLIP support natively in llama.cpp as a new architecture and in this case projects like llama-cpp-python can simply use the existing API.

[z3ugma]

@rlancemartin is that CLIP support natively in llama.cpp tracked under a specific issue or PR in the llama.cpp repo? cc @ggerganov

[ggerganov]

There was an attempt to implement a LLaVA API as part of the llama.cpp library here: ggerganov/llama.cpp#3613
But I don't really like the proposal, so I suggested to temporarily build a second library as part of the llava example until we support CLIP natively in llama.cpp (ggerganov/llama.cpp#3613 (comment))

Not sure if this approach (with the second llava lib) would work for llama-cpp-python though. On llama.cpp side, I think it would be rather easy to implement - just need to move some files around using the linked PR as a starting point and add build steps for the llava (or llava.cpp) lib. And then I guess llama-cpp-python would have to bind to both llama.cpp and llava libs

Other than this, we don't have a specific issue for tracking this yet. I think @monatis would be the best person to say what would it take to support CLIP arch straight into llama.cpp, as they implemented the current clip.cpp code.

[monatis]

If we decide to implement CLIP arch straight into llama.cpp, the required work depends on whether we want to support the full architecture or just as a image encoder for multimodal models. In any case, I need to revisit the conversion script and finalize the key-value pairs and tensor names to be future-proof.

For multimodal only:

1. Add image loading / preprocessing code. This will need linking stb-image.h to llama lib --not sure if we want this.
2. Update the current inference code to re-use the existing functionality in llama lib, e.g., model loading, offloading etc.

For the full architecture, additionally:
3. Implement CLIP tokenization, which is slightly different from the existing tokenizations in llama

These all can take some time. I initially started to implement in lmm.cpp, but I know from my clip.cpp experience it takes some time to actively support / manage a project on solo. I also have other ideas to combine all embedding models (CLIP, BERT, E5 etc.) in a embeddings library and then implement retrieval-augmented generation directly in C++ with this new lib plus llama. But (1) not sure where's the best place to implement it (as an example in llama.cpp or a downstream project etc.), and (2) maintenance burden.

Conclusion: My suggestion is to go on with @ggerganov's suggestion to build a llava lib as a part of examples in llama.cpp as a temporary solution but gradually (and hopefully quickly) move towards the full embeddings / multimodal / RAG experience as I can see that the community is more interested in such use cases than I thought in the first place. However, if this option does not work for your case, I'm also willing to implement it in such a way that you can link to. Community interests are what matters.

[abetlen]

@ggerganov I like that idea, as long as I can link to another shared library it's all good.

This approach would also be a good path to support things like finetuning directly from the python bindings without bloating the llama.h api surface.

[monatis]

@abetlen ggerganov/llama.cpp#3613 is ready for testing --does it cover your use case?

[abetlen]

Just published v0.2.15 with multimodal support.

Thank you so much @damian0815 and @monatis!

[abc2cba]

It seems the lllama.cpp have supported offload clip model to GPU, will it be supported for llama_cpp_python?