README.md

---
license: apache-2.0
language:
- en
- zh
base_model:
- Qwen/Qwen2.5-7B-Instruct
- BlinkDL/rwkv-7-world
pipeline_tag: text-generation
library_name: transformers
---

<div align="center">
  <img src="./figures/banner.jpg" style="border-radius: 10px; width: 100%; height: 100%; object-fit: cover;  box-shadow: 10px 10px 20px rgba(0, 0, 0, 0.5); border: 2px solid white;" alt="ARWKV" />
</div>


  <h1 align="center">ARWKV🪿</h1>

<p align="center">
  <a href="https://arxiv.org/abs/2501.15570"><b>Paper Link</b>👁️</a>  |  <a href="https://github.com/yynil/RWKVInside"><b>Github</b>✅</a>
</p>

# ARWKV-7B-GATE-MLP (Preview 0.1)

<img src="./figures/architecture.png" alt="ARWKV Hybrid Architecture"  width="30%">

*Preview version with **RWKV-7** time mixing and Transformer MLP*

## 📌 Overview

**ALL YOU NEED IS RWKV**

This is an **early preview** of our 7B parameter RNN-based model, trained on 2k context length **(only stage-2 applied, without SFT or DPO)** through 3-stage knowledge distillation from Qwen2.5-7B-Instruct. While being a foundational version, it demonstrates:

- ✅ RWKV-7's efficient recurrence mechanism
- ✅ No self-attention, fully O(n)
- ✅ Constant VRAM usage
- ✅ Single-GPU trainability

**Roadmap Notice**: We will soon open-source different enhanced versions with:
- 🚀 16k+ context capability
- 🧮 Math-specific improvements
- 📚 RL enhanced reasoning model

## How to use
```shell
pip3 install --upgrade rwkv-fla transformers
```

```python
from transformers import AutoModelForCausalLM, AutoTokenizer


model = AutoModelForCausalLM.from_pretrained(
    "RWKV-Red-Team/ARWKV-7B-Preview-0.1",
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(
    "RWKV-Red-Team/ARWKV-7B-Preview-0.1"
)
```

## 🔑 Key Features
| Component | Specification | Note |
|-----------|---------------|------|
| Architecture | RWKV-7 TimeMix + SwiGLU | Hybrid design |
| Context Window | 2048 training CTX | *Preview limitation* |
| Training Tokens | 40M | Distillation-focused |
| Precision | FP16 inference recommended(16G Vram required) | 15%↑ vs BF16 |

## 🏗️ Architecture Highlights
### Core Modification Flow
```diff
Qwen2.5 Decoder Layer:
- Grouped Query Attention
+ RWKV-7 Time Mixing (Eq.3)
- RoPE Positional Encoding
+ State Recurrence
= Hybrid Layer Output
```