ChatGLM3-6B

💻 项目代码库 • 🐦 推特 • 📃 [GLM@ACL 22] [代码库] • 📃 [GLM-130B@ICLR 23] [代码库]

👋 加入我们的Slack和微信社群

📍体验更大规模的ChatGLM模型，请访问chatglm.cn

更新说明 (Modification)

• 优化示例代码，新增NPU加速支持；

• 调整依赖库配置；

• Updated examples with NPU support;

• Modified dependencies;

模型介绍 (Introduction)

ChatGLM3-6B 作为ChatGLM系列的最新一代开源模型，在延续前两代模型对话流畅、部署便捷等优势的同时，具备以下显著特性：

1. 更强大的基座模型： ChatGLM3-6B的基座模型ChatGLM3-6B-Base采用更丰富的训练数据、更充分的训练迭代和更科学的训练策略。在语义理解、数学推理、代码生成、知识问答等各类基准测试中，ChatGLM3-6B-Base展现了同规模预训练模型中的顶尖性能。
2. 更完善的功能体系： ChatGLM3-6B采用全新设计的提示词格式，不仅支持流畅的多轮对话，还原生集成工具调用、代码解释器(Code Interpreter)和智能体任务等复杂场景处理能力。
3. 更完整的开源生态： 除对话模型ChatGLM3-6B外，同步开源基座模型ChatGLM-6B-Base、长文本对话模型ChatGLM3-6B-32K。所有模型权重对学术研究完全开放，提交登记问卷后可免费商用。

ChatGLM3-6B is the newest open-source model in the ChatGLM series. While maintaining the advantages of smooth conversation and easy deployment from its predecessors, it introduces remarkable features:

1. More Powerful Base Model: The foundation model ChatGLM3-6B-Base is trained with more diverse data, sufficient iterations, and optimized strategies. It demonstrates top-tier performance among pre-trained models of similar scale across various benchmarks including semantics, mathematics, coding, and knowledge QA.
2. Enhanced Functional Capabilities: Featuring a redesigned prompt format, ChatGLM3-6B supports not only seamless multi-turn dialogues but also native function calling, code interpreter, and agent task handling for complex scenarios.
3. Comprehensive Open-Source Ecosystem: Alongside the conversational ChatGLM3-6B, we open-source the base model ChatGLM-6B-Base and long-context variant ChatGLM3-6B-32K. All weights are fully accessible for academic research and free for commercial use after submitting the registration form.

环境依赖 (Dependencies)

pip install protobuf transformers==4.30.2 cpm_kernels torch>=2.0 gradio mdtex2html sentencepiece accelerate openmind

代码调用 (Code Usage)

通过以下代码即可调用 ChatGLM3-6B 模型生成对话内容：

You can generate dialogue responses by executing the following code snippet with the ChatGLM3-6B model:

from openmind import is_torch_npu_available, AutoTokenizer, AutoModel

if is_torch_npu_available():
    device = "npu:0"
elif torch.cuda.is_available():
    device = "cuda:0"
else:
    device = "cpu"

tokenizer = AutoTokenizer.from_pretrained("PyTorch-NPU/chatglm3_6b", trust_remote_code=True)
model = AutoModel.from_pretrained("PyTorch-NPU/chatglm3_6b", trust_remote_code=True, device_map=device).half()
model = model.eval()
response, history = model.chat(tokenizer, "你好", history=[])
print(response)
response, history = model.chat(tokenizer, "晚上睡不着应该怎么办", history=history)
print(response)

关于更多使用指南，包括如何运行命令行与网页版演示程序，以及通过模型量化技术优化显存占用，请参阅我们的 Github 项目库。

使用许可

本代码仓库遵循 Apache-2.0 开源协议，ChatGLM3-6B 模型权重的使用需遵守 模型许可协议。

文献引用

若您认为我们的研究成果对您有所裨益，敬请引用以下论文。

@article{zeng2022glm,
  title={Glm-130b: An open bilingual pre-trained model},
  author={Zeng, Aohan and Liu, Xiao and Du, Zhengxiao and Wang, Zihan and Lai, Hanyu and Ding, Ming and Yang, Zhuoyi and Xu, Yifan and Zheng, Wendi and Xia, Xiao and others},
  journal={arXiv preprint arXiv:2210.02414},
  year={2022}
}

@inproceedings{du2022glm,
  title={GLM: General Language Model Pretraining with Autoregressive Blank Infilling},
  author={Du, Zhengxiao and Qian, Yujie and Liu, Xiao and Ding, Ming and Qiu, Jiezhong and Yang, Zhilin and Tang, Jie},
  booktitle={Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={320--335},
  year={2022}
}