XVERSE-MoE-A36B
更新信息
- [2024/09/13] 发布 MoE 架构的 XVERSE-MoE-A36B 底座模型，Chat 对齐模型将在后续发布。
Update Information
- [2024/09/13] Released XVERSE-MoE-A36B MoE base model, the Chat version model will be released later.
模型介绍
XVERSE-MoE-A36B 是由深圳元象科技自主研发的支持多语言的大语言模型（Large Language Model），使用混合专家模型（MoE，Mixture-of-experts）架构，模型的总参数规模为 2554 亿，实际激活的参数量为 360 亿，本次开源的模型为底座模型 XVERSE-MoE-A36B，主要特点如下：
- 模型结构：XVERSE-MoE-A36B 为 Decoder-only 的 Transformer 架构，将密集模型的 FFN 层扩展为专家层，不同于传统 MoE 中每个专家的大小与标准 FFN 相同（如Mixtral 8x7B ），使用了更细粒度的专家，每个专家是标准 FFN 大小的 1/4，并设置了共享专家（Shared Expert）和非共享专家（Non-shared Expert）两类，共享专家在计算时始终被激活，非共享专家通过 Router 选择性激活。
- 训练数据：构建了海量高质量、多样化的数据对模型进行充分训练，包含中、英、俄、西等 40 多种语言，通过精细化设置不同类型数据的采样比例，使得中英两种语言表现优异，也能兼顾其他语言效果；模型使用 8K 长度的训练样本进行训练；在模型训练过程中进行了若干次数据的切换，来动态的引入持续处理的高质量数据，同时伴随数据采样比的调整。
- 训练策略：在切换数据的同时，为了使模型对新进数据进行快速且充分的学习，对学习率调度器也进行了相应调整。
- 训练框架：针对 MoE 模型中独有的专家路由和权重计算逻辑，进行了深入定制优化，开发出一套高效的融合算子，以提升计算效率。同时，为解决 MoE 模型显存占用和通信量大的挑战，设计了计算、通信和 CPU-Offload 的 Overlap 处理方式，从而提高整体吞吐量。
XVERSE-MoE-A36B 的模型大小、架构和学习率如下：
| total params | activated params | n_layers | d_model | n_heads | d_ff | n_non_shared_experts | n_shared_experts | top_k | lr |
|---|---|---|---|---|---|---|---|---|---|
| 255.4B | 36.5B | 50 | 6144 | 48 | 4096 | 64 | 2 | 6 | 2.5e−4 |
Model Introduction
XVERSE-MoE-A36B is a multilingual large language model, independently developed by Shenzhen Yuanxiang Technology which is using Mixture-of-experts (MoE) architecture. The total parameter scale of the model is 255 billion, with an actual number of activated parameters being 36 billion. The models released this time is the base model XVERSE-MoE-A36B. Its key features are as follows:
- Model Structure: XVERSE-MoE-A36B uses the mainstream Decoder-only Transformer network structure that extends the FFN layer of dense models to expert layers. Unlike traditional MoE model where each expert has the same size as standard FFN (such as Mixtral 8x7B), it uses more fine-grained experts, with each expert being 1/4 the size of a standard FFN. It includes shared experts and non-shared experts, where shared experts are always activated during computation, and non-shared experts are selectively activated through a Router.
- Training Data: The model has been thoroughly trained on a large-scale high-quality dataset, including more than 40 languages such as Chinese, English, Russian, and Spanish. The sampling ratio of different types of data is finely set, which makes the performance of Chinese and English excellent, and also takes into account the effect of other languages; The model is trained using training samples of length 8k; During the model training process, several data switches were made to dynamically introduce continuously processed high-quality data, along with adjustments to the data sampling ratio.
- Training Strategy: While switching data, corresponding adjustments were also made to the learning rate scheduler to ensure the model could quickly and thoroughly learn from the newly introduced data.
- Training Framework: We conducted in-depth customized optimization for the unique expert routing and weight calculation logic in the MoE model, developed an efficient fusion operator to improve computational efficiency. At the same time, to address the challenges of high memory consumption and communication volume in the MoE model, we designed a processing method for overlapping computation, communication, and CPU-Offload to increase overall throughput.
The models sizes, architectures and learning rate of XVERSE-MoE-A36B are showed as follows:
| total params | activated params | n_layers | d_model | n_heads | d_ff | n_non_shared_experts | n_shared_experts | top_k | lr |
|---|---|---|---|---|---|---|---|---|---|
| 255.4B | 36.5B | 50 | 6144 | 48 | 4096 | 64 | 2 | 6 | 2.5e−4 |
评测结果
为了综合评估模型的性能，我们在一系列标准数据集上进行了全面测试，包括MMLU、C-Eval、CMMLU、RACE-M、PIQA、GSM8K、MATH、MBPP和HumanEval，这些评估数据集覆盖了模型在多个领域的能力。并与相近参数规模的开源MoE模型进行了对比，结果如下：
对比开源 Base 模型 - MoE
| XVERSE-MoE-A36B | Grok-1-A85B | DeepSeek-V2-A21B | Skywork-MoE-A22B | Mixtral-8x22B-A39B | DBRX-A36B | |
|---|---|---|---|---|---|---|
| Total Params | 255B | 314B | 236B | 146B | 141B | 132B |
| MMLU | 80.8 | 73 | 78.5 | 77.4 | 77.8 | 73.7 |
| C-Eval | 79.5 | - | 81.7 | 82.2 | 56.8 | 44.9 |
| CMMLU | 81.7 | - | 84 | 79.5 | 59.9 | 61.3 |
| GSM8K | 89.5 | 62.9 | 79.2 | 76.1 | 82.3 | 70.7 |
| MATH | 53.3 | 23.9 | 43.6 | 31.9 | 34.1 | 25.6 |
| HumanEval | 51.8 | 63.2 | 48.8 | 43.9 | 45.1 | 46.3 |
| MBPP | 59.8 | - | 66.6 | - | 71.2 | 58 |
| PIQA | 84.8 | - | 83.7 | - | 84.1 | 84.5 |
| RACE-M | 88.4 | - | 73.1 | - | 85.7 | 55.9 |
对比开源 Base 模型 - Dense
| XVERSE-MoE-A36B | XVERSE-65B-2 | Llama3.1-405B | Nemotron-4-340B | Qwen1.5-110B | Qwen2-72B | Qwen1.5-72B | Llama3.1-70B | |
|---|---|---|---|---|---|---|---|---|
| Total Params | 255B | 65B | 405B | 340B | 110B | 72B | 72B | 70B |
| MMLU | 80.8 | 74.4 | 85.2 | 81.1 | 80.4 | 84.2 | 77.5 | 79.3 |
| C-Eval | 79.5 | 72.4 | - | - | 89.1 | 91 | 84.1 | - |
| CMMLU | 81.7 | 75.1 | - | - | 88.3 | 90.1 | 83.5 | - |
| GSM8K | 89.5 | 72.6 | 89 | - | 85.4 | 89.5 | 79.5 | 83.7 |
| MATH | 53.3 | 20.8 | 53.8 | - | 49.6 | 51.1 | 34.1 | 41.4 |
| HumanEval | 51.8 | 37.8 | 61 | 57.3 | 54.3 | 64.6 | 46.3 | 58.5 |
| MBPP | 59.8 | 40.6 | 73.4 | - | 70.9 | 76.9 | 66.9 | 66.2 |
| PIQA | 84.8 | 79.4 | 85.6 | - | - | - | 83.8 | |
| RACE-M | 88.4 | 90.7 | - | - | - | - | - |
对比闭源 Chat 模型
| XVERSE-MoE-A36B | GPT-4o | abab-6.5-20240415 | Step-2 | Baichuan3 | GLM-4 (0520) | |
|---|---|---|---|---|---|---|
| Total Params | 255B | - | 万亿 | 万亿 | 千亿 | - |
| MMLU | 80.8 | 88.7 | 78.7 | 81.7 | 83.3 | |
| C-Eval | 79.5 | - | - | - | - | - |
| CMMLU | 81.7 | - | - | - | 78.1 | - |
| GSM8K | 89.5 | - | 91.7 | 94 | 88.2 | 93.3 |
| MATH | 53.3 | 76.6 | 51.3 | 68.4 | 49.2 | 61.3 |
| HumanEval | 51.8 | 90.2 | 78 | 84.1 | 70.1 | 78.5 |
| MBPP | 59.8 | - | - | - | 68.2 | - |
| PIQA | 84.8 | - | - | - | - | - |
| RACE-M | 88.4 | - | - | - | - | - |
对于上述所有比较模型，我们汇报其官方结果与自测结果之间的最大值。
Model Evaluation
To comprehensively assess the performance of the model, we conducted extensive testing across a range of standard datasets, including MMLU, C-Eval, CMMLU, RACE-M, PIQA, GSM8K, Math, MBPP and HumanEval. And compared it with open-source MoE models of similar parameter scale, the results are as follows:
Comparison of Open-Weight Base Models - MoE
| XVERSE-MoE-A36B | Grok-1-A85B | DeepSeek-V2-A21B | Skywork-MoE-A22B | Mixtral-8x22B-A39B | DBRX-A36B | |
|---|---|---|---|---|---|---|
| Total Params | 255B | 314B | 236B | 146B | 141B | 132B |
| MMLU | 80.8 | 73 | 78.5 | 77.4 | 77.8 | 73.7 |
| C-Eval | 79.5 | - | 81.7 | 82.2 | 56.8 | 44.9 |
| CMMLU | 81.7 | - | 84 | 79.5 | 59.9 | 61.3 |
| GSM8K | 89.5 | 62.9 | 79.2 | 76.1 | 82.3 | 70.7 |
| MATH | 53.3 | 23.9 | 43.6 | 31.9 | 34.1 | 25.6 |
| HumanEval | 51.8 | 63.2 | 48.8 | 43.9 | 45.1 | 46.3 |
| MBPP | 59.8 | - | 66.6 | - | 71.2 | 58 |
| PIQA | 84.8 | - | 83.7 | - | 84.1 | 84.5 |
| RACE-M | 88.4 | - | 73.1 | - | 85.7 | 55.9 |
Comparison of Open-Weight Base Models - Dense
| XVERSE-MoE-A36B | XVERSE-65B-2 | Llama3.1-405B | Nemotron-4-340B | Qwen1.5-110B | Qwen2-72B | Qwen1.5-72B | Llama3.1-70B | |
|---|---|---|---|---|---|---|---|---|
| Total Params | 255B | 65B | 405B | 340B | 110B | 72B | 72B | 70B |
| MMLU | 80.8 | 74.4 | 85.2 | 81.1 | 80.4 | 84.2 | 77.5 | 79.3 |
| C-Eval | 79.5 | 72.4 | - | - | 89.1 | 91 | 84.1 | - |
| CMMLU | 81.7 | 75.1 | - | - | 88.3 | 90.1 | 83.5 | - |
| GSM8K | 89.5 | 72.6 | 89 | - | 85.4 | 89.5 | 79.5 | 83.7 |
| MATH | 53.3 | 20.8 | 53.8 | - | 49.6 | 51.1 | 34.1 | 41.4 |
| HumanEval | 51.8 | 37.8 | 61 | 57.3 | 54.3 | 64.6 | 46.3 | 58.5 |
| MBPP | 59.8 | 40.6 | 73.4 | - | 70.9 | 76.9 | 66.9 | 66.2 |
| PIQA | 84.8 | 79.4 | 85.6 | - | - | - | 83.8 | |
| RACE-M | 88.4 | 90.7 | - | - | - | - | - |
Comparison of Closed-Source Chat Models
| XVERSE-MoE-A36B | GPT-4o | abab-6.5-20240415 | Step-2 | Baichuan3 | GLM-4 (0520) | |
|---|---|---|---|---|---|---|
| Total Params | 255B | - | Trillion scale | Trillion scale | Hundred billion scale | - |
| MMLU | 80.8 | 88.7 | 78.7 | 81.7 | 83.3 | |
| C-Eval | 79.5 | - | - | - | - | - |
| CMMLU | 81.7 | - | - | - | 78.1 | - |
| GSM8K | 89.5 | - | 91.7 | 94 | 88.2 | 93.3 |
| MATH | 53.3 | 76.6 | 51.3 | 68.4 | 49.2 | 61.3 |
| HumanEval | 51.8 | 90.2 | 78 | 84.1 | 70.1 | 78.5 |
| MBPP | 59.8 | - | - | - | 68.2 | - |
| PIQA | 84.8 | - | - | - | - | - |
| RACE-M | 88.4 | - | - | - | - | - |
For all the comparison models mentioned above, we report the maximum value between their official results and our self-evaluation results.
使用方法
Transformers 加载方式
可通过以下代码加载 XVERSE-MoE-A36B 模型来进行推理：
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("xverse/XVERSE-MoE-A36B")
model = AutoModelForCausalLM.from_pretrained("xverse/XVERSE-MoE-A36B", trust_remote_code=True, torch_dtype=torch.bfloat16, device_map='auto')
model = model.eval()
inputs = tokenizer('北京的景点：故宫、天坛、万里长城等。\n深圳的景点：', return_tensors='pt').input_ids
inputs = inputs.cuda()
generated_ids = model.generate(inputs, max_new_tokens=70, eos_token_id=tokenizer.eos_token_id, repetition_penalty=1.1)
print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True))
Usage
Loading with Transformers
The XVERSE-MoE-A36B model can be loaded for inference using the following code:
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("xverse/XVERSE-MoE-A36B")
model = AutoModelForCausalLM.from_pretrained("xverse/XVERSE-MoE-A36B", trust_remote_code=True, torch_dtype=torch.bfloat16, device_map='auto')
model = model.eval()
inputs = tokenizer('北京的景点：故宫、天坛、万里长城等。\n深圳的景点：', return_tensors='pt').input_ids
inputs = inputs.cuda()
generated_ids = model.generate(inputs, max_new_tokens=70, eos_token_id=tokenizer.eos_token_id, repetition_penalty=1.1)
print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True))
局限性与免责申明
XVERSE-MoE-A36B 与其他所有 LLM 一样，在某些情况下可能会产生不准确、有偏见或其他令人反感的内容。因此，请谨慎使用模型生成的内容，请勿将生成的有害内容进行传播，在部署任何 XVERSE-MoE-A36B 的应用之前，开发人员应根据其具体应用对模型进行安全测试和调优。
我们强烈警告不要将 XVERSE-MoE-A36B 模型用于制造或传播有害信息，或进行任何可能损害公众、国家、社会安全或违反法规的活动。如果使用 XVERSE-MoE-A36B 模型产生任何问题，无论是数据安全问题、公共舆论风险，还是模型被误解、滥用、传播或不合规使用所引发的任何风险和问题，我们将不承担任何责任。
模型开源协议
使用本仓库的源码需要遵循 Apache-2.0 开源协议，使用 XVERSE-MoE-A36B 的模型权重则需要遵循模型许可协议。
XVERSE-MoE-A36B 模型权重对学术研究完全开放，并且支持免费商用。如需申请商业许可证，请填写【申请表】，如有其他问题或合作，请联系 [email protected]。
Limitations and Disclaimer
Like all other Large Language Models (LLMs), XVERSE-MoE-A36B may produce inaccurate, biased, or otherwise offensive content under certain circumstances. Therefore, please use the content generated by the model with caution and refrain from disseminating harmful content. Before deploying any application of XVERSE-MoE-A36B, developers should conduct safety tests and optimization of the model according to its specific application.
We strongly warn against the use of the XVERSE-MoE-A36B model for producing or spreading harmful information, or conducting any activities that might harm the public, national, or social security, or violate regulations. We assume no responsibility for any problems arising from the use of the XVERSE-MoE-A36B model, whether it be data security issues, public opinion risks, or any risks and issues caused by misunderstanding, misuse, dissemination, or non-compliance with the model.
Open Source License
The use of the source code in this repository must follow the Apache-2.0 open-source license, while the use of the model weights of XVERSE-MoE-A36B needs to adhere to the Model License Agreement.
The XVERSE-MoE-A36B model weights are fully open to academic research and support free commercial use. To apply for a commercial license, please fill in the application form. For other questions or collaborations, please contact [email protected].
- Downloads last month
- 25