yujiepan/llama-3.2-vision-tiny-random

This model is for debugging. It is randomly initialized using the config from meta-llama/Llama-3.2-90B-Vision-Instruct but with smaller size.

Codes:

import os

import accelerate
import requests
import torch
import transformers
from huggingface_hub import create_repo, upload_folder
from PIL import Image
from transformers import AutoProcessor, MllamaForConditionalGeneration
from transformers.models.mllama import MllamaConfig

model_id = 'meta-llama/Llama-3.2-90B-Vision-Instruct'
repo_id = 'yujiepan/llama-3.2-vision-tiny-random'
save_path = f'/tmp/{repo_id}'

os.system(f'rm -rf {save_path}')

config = transformers.AutoConfig.from_pretrained(
    model_id,
    trust_remote_code=True,
)
config.text_config.hidden_size = 8
config.text_config.intermediate_size = 16
config.text_config.num_attention_heads = 2
config.text_config.num_key_value_heads = 1
config.text_config.num_hidden_layers = 2
config.text_config.cross_attention_layers = [1]

config.vision_config.attention_heads = 2
config.vision_config.hidden_size = 8
config.vision_config.intermediate_size = 16
config.vision_config.intermediate_layers_indices = [0]
config.vision_config.num_global_layers = 2
config.vision_config.num_hidden_layers = 2
config.vision_config.vision_output_dim = 16


transformers.set_seed(42)
model = MllamaForConditionalGeneration(config)
model.generation_config = transformers.GenerationConfig.from_pretrained(
    model_id)
model = model.to(torch.bfloat16)

transformers.set_seed(42)
with torch.no_grad():
    for p in model.parameters():
        torch.nn.init.normal_(p)

model.save_pretrained(save_path)

processor = AutoProcessor.from_pretrained(model_id)
processor.save_pretrained(save_path)

url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/0052a70beed5bf71b92610a43a52df6d286cd5f3/diffusers/rabbit.jpg"
image = Image.open(requests.get(url, stream=True).raw)

messages = [
    {"role": "user", "content": [
        {"type": "image"},
        {"type": "text", "text": "If I had to write a haiku for this one, it would be: "}
    ]}
]
input_text = processor.apply_chat_template(
    messages, add_generation_prompt=True)
inputs = processor(image, input_text, return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=30)
print(processor.decode(output[0]))

os.system(f'ls -alh {save_path}')
# os.system(f'rm -rf {save_path}/model.safetensors')
# create_repo(repo_id, exist_ok=True)
# upload_folder(repo_id=repo_id, folder_path=save_path)