Intel/llava-llama-3-8b

Model Details: LLaVA-llama-3-8B

llava-llama-3-8b is a large multimodal model (LMM) trained using the LLaVA-v1.5 framework with the 8-billion parameter meta-llama/Meta-Llama-3-8B-Instruct model as language backbone and the CLIP-based vision encoder.

Model Details	Description
Authors	Intel: Musashi Hinck*, Matthew L. Olson*, Vasudev Lal
Date	May 2024
Version	1
Type	Large multimodal model (LMM)
Paper or Other Resources	Improved Baselines with Visual Instruction Tuning
License	Intel Research Use License
Questions or Comments	Community Tab and Intel DevHub Discord

This model card was created by Eduardo Alvarez and the authors listed above.

Intended Use

Intended Use	Description
Primary intended uses	The model has been finetuned for multimodal benchmark evaluations, but can also be used as a multimodal chatbot.
Primary intended users	Anyone using or evaluating multimodal models.
Out-of-scope uses	This model is not intended for uses that require high levels of factuality, high stakes situations, mental health or medical applications, generating misinformation or disinformation, impersonating others, facilitating or inciting harassment or violence, any use that could lead to the violation of a human right under the UN Declaration of Human Rights.

How to use

Please note, we only provide the trained weights difference and do not provide a copy of the base meta-llama/Meta-Llama-3-8B-Instruct model. Any use of these weights requires a separate download of the base model.

# Copyright 2024 Intel Corporation
# SPDX-License-Identifier: Apache-2.0

import requests
import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForPreTraining
import transformers

def expand2square(pil_img, background_color):
    width, height = pil_img.size
    if width == height:
        return pil_img
    elif width > height:
        result = Image.new(pil_img.mode, (width, width), background_color)
        result.paste(pil_img, (0, (width - height) // 2))
        return result
    else:
        result = Image.new(pil_img.mode, (height, height), background_color)
        result.paste(pil_img, ((height - width) // 2, 0))
        return result

def add_model_a_to_b(model_a, model_b):
    state_dict_a = model_a.state_dict()
    state_dict_b = model_b.state_dict()

    # Ensure keys match before subtraction
    if set(state_dict_a.keys()) != set(state_dict_b.keys()):
        raise ValueError("Model state dicts do not have the same keys.")

    for key in state_dict_a:
        if state_dict_a[key].shape != state_dict_b[key].shape:
            raise ValueError(f"Shape mismatch for key '{key}': {state_dict_a[key].shape} vs {state_dict_b[key].shape}")
        # Subtract model_a's weights from model_b for the matching key
        state_dict_b[key] = state_dict_b[key] + state_dict_a[key]

    # Update model_b with the new weights
    model_b.load_state_dict(state_dict_b)

output_checkpoint = "" # set if you don't want to merge every time
hf_checkpoint = "Intel/llava-llama-3-8b"
device = "cuda" if torch.cuda.is_available() else "cpu"

processor = AutoProcessor.from_pretrained(hf_checkpoint)
model = AutoModelForPreTraining.from_pretrained(hf_checkpoint)
if model.language_model.model.embed_tokens.weight[-1].sum() == 0:
    print("adding llama3 weights")
    model_id = "meta-llama/Meta-Llama-3-8B-Instruct"
    pipeline = transformers.pipeline(
        "text-generation",
        model=model_id,
        model_kwargs={"torch_dtype": torch.bfloat16},
        device_map="cpu",
    )
    llama3 = pipeline.model
    add_model_a_to_b(llama3, model.language_model)
    if output_checkpoint:
        print("saving weights, so no adding is needed again")
        model.save_pretrained(output_checkpoint)

device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

prompt = processor.tokenizer.apply_chat_template(
    [{'role': 'user', 'content': "<image>\nWhat's the content of the image?"}],
    tokenize=False,
    add_generation_prompt=True
)

url = "https://www.ilankelman.org/stopsigns/australia.jpg"
image = Image.open(requests.get(url, stream=True).raw)

#original llava pads with mean, HF llava pads with zeros
image = expand2square(image, tuple(int(x*255) for x in processor.image_processor.image_mean)) 
inputs = processor(text=prompt, images=image, return_tensors="pt").to(device)
# Generate
generate_ids = model.generate(**inputs, max_length=30)
output = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
print(output)

Factors

Factors	Description
Environment	Trained on a 4 node cluster with a total of 32 Gaudi 2 accelerators
Card Prompts	Model training and deployment on alternate hardware and software will change model performance

Training Data

The model was trained using the LLaVA-v1.5 data mixture. This is listed as follows:

• 558K filtered image-text pairs from LAION/CC/SBU, captioned by BLIP.
• 158K GPT-generated multimodal instruction-following data.
• 450K academic-task-oriented VQA data mixture.
• 40K ShareGPT data.

Ethical Considerations

Intel is committed to respecting human rights and avoiding causing or contributing to adverse impacts on human rights. See Intel’s Global Human Rights Principles. Intel’s products and software are intended only to be used in applications that do not cause or contribute to adverse impacts on human rights.

Ethical Considerations	Description
Data	The model was trained using the LLaVA-v1.5 data mixture as described above.
Human life	The model is not intended to inform decisions central to human life or flourishing.
Mitigations	No additional risk mitigation strategies were considered during model development.
Risks and harms	This model has not been assessed for harm or biases, and should not be used for sensitive applications where it may cause harm.
Use cases	-

Caveats and Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. This model has not been assessed for harm or biases, and should not be used for sensitive applications where it may cause harm.