metadata

license: apache-2.0
language:
  - en
library_name: transformers
tags:
  - multimodal
  - aria

Aria Model Card

Key features

SoTA Multimodal Native Performance: Aria achieves strong performance on a wide range of multimodal, language, and coding tasks. It is superior in video and document understanding.
Lightweight and Fast: Aria is a mixture-of-expert model with 3.9B activated parameters per token. It efficently encodes visual input of variable sizes and aspect ratios.
Long Multimodal Context Window: Aria supports multimodal input of up to 64K tokens. It can caption a 256-frame video in 10 seconds.

Benchmark

Category	Benchmark	Aria	Pixtral 12B	Llama3.2 11B	GPT-4o mini	GPT-4o	Gemini-1.5 Flash	Gemini-1.5 Pro
Knowledge (Multimodal)	MMMU	54.9	52.5	49.6	59.4	69.1	56.1	62.2
Math (Multimodal)	MathVista	66.1	58.0	51.5	-	54.7	63.8	58.4
Document	DocQA	92.6	90.7	84.4	-	92.8	89.9	93.1
Chart	ChartQA	86.4	81.8	78.7	-	85.7	85.4	87.2
Scene Text	TextVQA	81.1	-	78.2	-	-	78.7	78.7
General Visual QA	MMBench-1.1	80.3	-	-	76.0	82.2	-	73.9
Video Understanding	LongVideoBench	66.6	47.4	45.7	58.8	66.7	62.4	64.4
Knowledge (Language)	MMLU (5-shot)	73.3	69.2	69.4	-	89.1	78.9	85.9
Math (Language)	MATH	50.8	48.1	51.9	70.2	76.6	-	-
Reasoning (Language)	ARC Challenge	91.0	-	83.4	96.4	96.7	-	-
Coding	HumanEval	73.2	72.0	72.6	87.2	90.2	74.3	84.1

Quick Start

Installation

pip install git+github.com/rhymes-ai/Aria.git
pip install flash-attn --no-build-isolation

Inference

Aria has 25.3B total parameters, it can be loaded in one A100 (80GB) GPU with bfloat16 precision.

Here is a code snippet to show you how to use Aria.

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

model_id_or_path = "rhymes-ai/Aria"

model = AutoModelForCausalLM.from_pretrained(model_id_or_path, device_map="auto", torch_dtype=torch.bfloat16, trust_remote_code=True)

processor = AutoProcessor.from_pretrained(model_id_or_path, trust_remote_code=True)

image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png"

image = Image.open(requests.get(image_path, stream=True).raw)

messages = [
    {
        "role": "user",
        "content": [
            {"text": None, "type": "image"},
            {"text": "what is the image?", "type": "text"},
        ],
    }
]

text = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(text=text, images=image, return_tensors="pt")
inputs["pixel_values"] = inputs["pixel_values"].to(model.dtype)
inputs = {k: v.to(model.device) for k, v in inputs.items()}

with torch.inference_mode(), torch.cuda.amp.autocast(dtype=torch.bfloat16):
    output = model.generate(
        **inputs,
        max_new_tokens=500,
        stop_strings=["<|im_end|>"],
        tokenizer=processor.tokenizer,
        do_sample=True,
        temperature=0.9,
    )
    output_ids = output[0][inputs["input_ids"].shape[1]:]
    result = processor.decode(output_ids, skip_special_tokens=True)

print(result)

Advanced Inference and Fine-tuning

We provide a codebase for more advanced usage of Aria, including vllm inference, cookbooks, and fine-tuning on custom datasets.

Citation

If you find our work helpful, please consider citing.

@article{aria,
  title={},
  author={},
  year={2024},
  journal={}
}