license: apache-2.0
Yuan2.0-M32: Mixture of Experts with Attention Router
👾 ModelScope • 🤗 Hugging Face • 💬 WeChat• 📎 Yuan2.0-M32 Paper
1. Introduction
Yuan2.0-M32 is a Mixture-of-Experts (MoE) language model with 32 experts, of which 2 are active. A new router network, Attention Router, is proposed and has been adopted for more efficient expert selection, boosting accuracy by 3.8% over models using a classical router network. Yuan 2.0-M32 is trained from scratch with 2000B tokens, and its training computation is only 9.25% of that required by a dense model of the same parameter scale. Demonstrating competitive capabilities in coding, math, and various specialized fields, Yuan2.0-M32 operates with only 3.7B active parameters out of a total 40B, and a forward computation of 7.4 GFLOPS per token, which is just 1/19th of Llama3-70B's requirement. Yuan 2.0-M32 has surpassed Llama3-70B on the MATH and ARC-Challenge benchmarks, achieving accuracies of 55.9% and 95.8%, respectively. The basic information of the Yuan2.0-M32 model is as follows:
- Total Parameters ： 40B
- Experts： 32
- Active Experts： 2
- Active Parameters： 3.7B
- Training Tokens： 2000B tokens
- Sequence Length： 16K
The technical report for the Yuan2.0-M32 model has been released, and you can find more detailed technical information and evaluation results by referring to the paper.
2. Model Downloads
| Model | Sequence Length | Type | Download |
|---|---|---|---|
| Yuan2.0-M32 | 16K | Megatron | HuggingFace |
| Yuan2.0-M32-HF | 16K | HuggingFace | HuggingFace |
| Yuan2.0-M32-GGUF | 16K | GGUF | HuggingFace |
| Yuan2.0-M32-GGUF-INT4 | 16K | GGUF | HuggingFace |
3. Evaluation
3.1 Benchmarks 🏆
We conducted a thorough evaluation of the Yuan2.0-M32 model across a range of benchmarks, including HumanEval, GSM8K, MMLU, Math, and ARC-Challenge. These benchmarks are designed to test the model's proficiency in key areas such as natural language understanding, knowledge acquisition, mathematical computation and reasoning, and code generation. The Yuan2.0-M32 has shown a consistent and significant advantage over other models like Llama3-8B and Mistral-8×7B, excelling in all evaluated tasks. Remarkably, its overall performance is on par with the more substantial Llama3-70B model.The detailed evaluation results are outlined in the subsequent table.
| Model | HumanEval | GSM8K | MMLU | Math | ARC-C* |
|---|---|---|---|---|---|
| Llama3-70B | 81.7% | 93% | 80.3 | 50.4% | 93.3% |
| Llama3-8B | 62.2% | 79.6% | 68.4% | 30% | 78.6% |
| Phi-3-medium | 62.2% | 91.0% | 78.0% | - | 91.6% |
| Phi-3-small | 61% | 89.6% | 75.7% | - | 90.7% |
| Phi-3-mini | 58.5% | 82.5% | 68.8% | - | 84.9% |
| Mistral-8*22B | 45.1% | 78.6% | 77.8% | 41,8% | 91.3% |
| Mistral-8*7B | 40.2% | 58.4% | 70.86% | 28.4% | 85.9% |
| Yuan2.0-M32 | 74.4% | 92.7% | 72.2% | 55.9% | 95.8% |
* ARC-C: AI2 Reasoning Challenge (ARC) benchmark contains more complex parts that need further reasoning.
3.2 Computational Utilization for Model
| Model | Params (B) | Active Params (B) | GFLOPs/token (Inference) | GFLOPS/token (Fine-tune) | Mean Accuracy | Average Accuracy/GFLOPSs per token (Inference) |
|---|---|---|---|---|---|---|
| Llama3-70B | 70 | 70 | 140 | 420 | 79.25 | 0.57 |
| Llama3-8B | 8 | 8 | 16 | 48 | 64.15 | 4.00 |
| Mistral-8*22B | 141 | 39 | 78 | 234 | 72.38 | 0.93 |
| Mistral-8*7B | 47 | 12.9 | 25.8 | 77.3 | 60.83 | 2.36 |
| Yuan2.0-M32 | 40 | 3.7 | 7.4 | 22.2 | 79.15 | 10.69 |
4. Quick Start
4.1 Environment Config
We strongly recommend using the latest release of docker images of Yuan2.0-M32.You can launch an instance of the Yuan 2.0 container with the following Docker commands:
docker pull yuanmodel/yuan2.0:m32
docker run --gpus all --privileged --ulimit stack=68719476736 --shm-size=1000G -itd -v /path/to/yuan_2.0:/workspace/yuan_2.0 -v /path/to/dataset:/workspace/dataset -v /path/to/checkpoints:/workspace/checkpoints --name your_name yuanmodel/yuan2.0:m32
docker exec -it your_name bash
4.2 Data Preprocess
We have provided the data preprocess script. See documentation here.
4.3 Model Pretrain
We've provided several scripts for pretraining in the example. The details can be seen from documentation here.
4.4 Inference Service
For a detailed deployment plan, please refer to vllm.
5. Statement of Agreement
The use of the source code in this repository requires compliance with the open source license agreement Apache 2.0. The Yuan2.0 model supports commercial use and does not require authorization. Please understand and comply with the 《Yuan2.0 Model License Agreement》. Do not use the open source model and code, as well as derivatives generated from open source projects, for any purposes that may cause harm to the country and society, or for any services that have not undergone security assessment and filing. Although we have taken measures to ensure the compliance and accuracy of the data during training, the model has a huge number of parameters and is affected by probability and randomness factors. We cannot guarantee the accuracy of the output content, and the model is easily misled by input instructions. This project does not assume any data security, public opinion risks, or any model misleading, abusing, spreading caused by open-source models and code Risks and responsibilities arising from improper utilization You will be solely responsible for the risks and consequences arising from the use, copying, distribution, and modification of the model in this open source project
6. Contact Us
If you have any questions, please raise an issue or contact us at [email protected]