Fix support for SGLang inference

#2
No description provided.

+1

Would it be possible to add the processor configs to the other llava 1.6 models after this is confirmed to be working?

+1

Would it be possible to add the processor configs to the other llava 1.6 models after this is confirmed to be working?

I have already confirmed this to be working as I came up with this after hours of trying and merging fixes from different sources

I have some error when running it
KeyError: 'model.image_newline'

image.png

Fix: Worked with sglang==0.1.9

I got it working with sglang==0.1.11 and 0.1.12 too

Can we get it merged? I would like to use this model in Google Colab.

Since this seems stalled, I made equivalent changes in a fork of this repo: https://huggingface.co/SurfaceData/llava-v1.6-mistral-7b-sglang

@fozziethebeat I tried to convert your model but it also failed;

Traceback (most recent call last):
  File "/content/drive/MyDrive/AI/llama.cpp/convert.py", line 1486, in <module>
    main()
  File "/content/drive/MyDrive/AI/llama.cpp/convert.py", line 1472, in main
    model   = convert_model_names(model, params, args.skip_unknown)
  File "/content/drive/MyDrive/AI/llama.cpp/convert.py", line 1217, in convert_model_names
    raise Exception(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
Exception: Unexpected tensor name: model.image_newline. Use --skip-unknown to ignore it (e.g. LLaVA)

What is this conversion script for? I only prepared the model to work for SGLang, so can't make any promises it'll work for other types of conversion.

it's the script coming with llama.cpp to convert a pytorch model to GGUF. I get exactly the same error, so I thought it's the same problem but probably it's not. https://github.com/ggerganov/llama.cpp/blob/master/convert.py

This PR is not meant for llama cpp and it is only a coincidence if it work with llama cpp. This is strictly for SGLang only.

@olafgeibig llama.cpp has a detailed explanation on how to convert llava and llava1.6 models to gguf format.
https://github.com/ggerganov/llama.cpp/tree/master/examples/llava

Also no need to convert it since there are already pre converted llava 1.6 models. Just search llava 1.6 gguf in huggingface

Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment