Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,132 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
language:
|
4 |
+
- en
|
5 |
+
library_name: transformers
|
6 |
+
pipeline_tag: image-text-to-text
|
7 |
+
tags:
|
8 |
+
- multimodal
|
9 |
+
- aria
|
10 |
+
---
|
11 |
+
<!-- <p align="center">
|
12 |
+
<br>Aria</br>
|
13 |
+
</p> -->
|
14 |
+
|
15 |
+
|
16 |
+
# Aria-Base-64K Model Card
|
17 |
+
|
18 |
+
<p align="center">
|
19 |
+
🔗 <a href="https://rhymes.ai/" target="_blank"> Try Aria!</a> · 📖 <a href="https://www.rhymes.ai/blog-details/aria-first-open-multimodal-native-moe-model" target="_blank">Blog</a> · 📌 <a href="https://arxiv.org/pdf/2410.05993" target="_blank">Paper</a>
|
20 |
+
· ⭐ <a href="https://github.com/rhymes-ai/Aria" target="_blank">GitHub</a> · 🟣 <a href="https://discord.com/invite/u8HxU23myj" target="_blank"> Discord </a>
|
21 |
+
</p>
|
22 |
+
|
23 |
+
|
24 |
+
This checkpoint is one of base models of [Aria](https://huggingface.co/rhymes-ai/Aria), designed for research purposes as well as continue training. Specifically, Aria-Base-64K corresponds to the model checkpoint after the long-context pre-training stage (boxed in purple).
|
25 |
+
|
26 |
+
<img src="./aria-stages.png" alt="Aria Training Stages" style="width: 75%;">
|
27 |
+
|
28 |
+
Aria-Base-64K is fine-tuned from [Aria-Base-8K](https://huggingface.co/teowu/Aria-Base-8K).
|
29 |
+
|
30 |
+
<!--
|
31 |
+
- Aria is the **first open multimodal native MoE** model, capable of seamlessly handling various input modalities within a MoE architecture.
|
32 |
+
- Aria performs **on par with GPT-4o mini and Gemini 1.5 Flash** across a range of multimodal tasks while maintaining strong performance on **text**-only tasks.
|
33 |
+
- Compared to similar or even larger models, Aria boasts **faster speeds** and **lower costs**. This high efficiency stems from its ability to activate only 3.9B parameters during inference – the **fewest** among models with comparable performance.
|
34 |
+
-->
|
35 |
+
|
36 |
+
## Aria-Base-8K
|
37 |
+
|
38 |
+
- **Base Model After Long-Context Pre-training**: This model corresponds to the model checkpoint after the long-context pre-training stage, with 33B tokens (21B multimodal, 12B language, 69% in long-form) trained in this stage. This stage lasts 1,000 iterations, with all sequences packed to 65536 with Megatron-LM, with global batch size 512. During this training stage, the learning rate keeps constant at `3.5e-5`.
|
39 |
+
- **Appropriate for Video and Long-document Fine-tuning**: This model is recommended for long-form continue pre-training or fine-tuning, e.g. on video QA datasets or long-document QA datasets. While resource is limited, it is also possible to post-train this model with short instruction tuning datasets and transfer to long-form QA scenarios.
|
40 |
+
- **Understanding on Hundreds of Images**: This model is capable of understanding up to 250 high-resolution images or up to 500 mid-resolution images.
|
41 |
+
- **Strong Base Performance on Language and Multimodal Scenarios**: This model retains strong base performance as [Aria-Base-8K](https://huggingface.co/teowu/Aria-Base-8K).
|
42 |
+
- ***Limited Chat Template Availability***: This model is trained with a very low percentage of data (around 3%) re-formatted with the chat template. Hence, it might not be optimal to be directly tested with various benchmarks.
|
43 |
+
|
44 |
+
<!-- # Model Info
|
45 |
+
|
46 |
+
| Model | Download | Parameter | Context Length |
|
47 |
+
| :---- | :------- | :------------ | :------ |
|
48 |
+
| Aria | < HF link - TBD> | • Activation: 3.9B (3.5B MoE + 0.4B Visual Encoder) <br> • Total: 25.3B | 64K | -->
|
49 |
+
|
50 |
+
## Benchmark
|
51 |
+
|
52 |
+
N/A.
|
53 |
+
|
54 |
+
## Quick Start
|
55 |
+
### Installation
|
56 |
+
```
|
57 |
+
pip install transformers==4.45.0 accelerate==0.34.1 sentencepiece==0.2.0 torchvision requests torch Pillow
|
58 |
+
pip install flash-attn --no-build-isolation
|
59 |
+
|
60 |
+
# For better inference performance, you can install grouped-gemm, which may take 3-5 minutes to install
|
61 |
+
pip install grouped_gemm==0.1.6
|
62 |
+
```
|
63 |
+
|
64 |
+
### Inference
|
65 |
+
|
66 |
+
You can use the same method as the final Aria model to load this checkpoint. However, as the base model, it might not be able to yield optimal chat performance.
|
67 |
+
|
68 |
+
```python
|
69 |
+
import requests
|
70 |
+
import torch
|
71 |
+
from PIL import Image
|
72 |
+
from transformers import AutoModelForCausalLM, AutoProcessor
|
73 |
+
|
74 |
+
model_id_or_path = "teowu/Aria-Base-64K"
|
75 |
+
|
76 |
+
model = AutoModelForCausalLM.from_pretrained(model_id_or_path, device_map="auto", torch_dtype=torch.bfloat16, trust_remote_code=True)
|
77 |
+
|
78 |
+
processor = AutoProcessor.from_pretrained(model_id_or_path, trust_remote_code=True)
|
79 |
+
|
80 |
+
image_path = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cat.png"
|
81 |
+
|
82 |
+
image = Image.open(requests.get(image_path, stream=True).raw)
|
83 |
+
|
84 |
+
messages = [
|
85 |
+
{
|
86 |
+
"role": "user",
|
87 |
+
"content": [
|
88 |
+
{"text": None, "type": "image"},
|
89 |
+
{"text": "what is the image?", "type": "text"},
|
90 |
+
],
|
91 |
+
}
|
92 |
+
]
|
93 |
+
|
94 |
+
text = processor.apply_chat_template(messages, add_generation_prompt=True)
|
95 |
+
inputs = processor(text=text, images=image, return_tensors="pt")
|
96 |
+
inputs["pixel_values"] = inputs["pixel_values"].to(model.dtype)
|
97 |
+
inputs = {k: v.to(model.device) for k, v in inputs.items()}
|
98 |
+
|
99 |
+
with torch.inference_mode(), torch.cuda.amp.autocast(dtype=torch.bfloat16):
|
100 |
+
output = model.generate(
|
101 |
+
**inputs,
|
102 |
+
max_new_tokens=500,
|
103 |
+
stop_strings=["<|im_end|>"],
|
104 |
+
tokenizer=processor.tokenizer,
|
105 |
+
do_sample=True,
|
106 |
+
temperature=0.9,
|
107 |
+
)
|
108 |
+
output_ids = output[0][inputs["input_ids"].shape[1]:]
|
109 |
+
result = processor.decode(output_ids, skip_special_tokens=True)
|
110 |
+
|
111 |
+
print(result)
|
112 |
+
```
|
113 |
+
|
114 |
+
### Advanced Inference and Fine-tuning
|
115 |
+
|
116 |
+
We provide a [codebase](https://github.com/rhymes-ai/Aria) for more advanced usage of Aria,
|
117 |
+
including vllm inference, cookbooks, and fine-tuning on custom datasets.
|
118 |
+
|
119 |
+
As it shares the same structure with the final model,
|
120 |
+
you may just replace the `rhymes-ai/Aria` to this model path for any advanced inference and fine-tuning.
|
121 |
+
|
122 |
+
|
123 |
+
## Citation
|
124 |
+
If you find our work helpful, please consider citing.
|
125 |
+
```
|
126 |
+
@article{aria,
|
127 |
+
title={Aria: An Open Multimodal Native Mixture-of-Experts Model},
|
128 |
+
author={Dongxu Li and Yudong Liu and Haoning Wu and Yue Wang and Zhiqi Shen and Bowen Qu and Xinyao Niu and Guoyin Wang and Bei Chen and Junnan Li},
|
129 |
+
year={2024},
|
130 |
+
journal={arXiv preprint arXiv:2410.05993},
|
131 |
+
}
|
132 |
+
```
|