ChatGLM-6B is an open bilingual language model based on [General Language Model (GLM)](https://github.com/THUDM/GLM) framework, with 6.2 billion parameters. With the quantization technique, users can deploy locally on consumer-grade graphics cards (only 6GB of GPU memory is required at the INT4 quantization level). Welcome to use the larger ChatGLM model on [chatglm.cn](https://chatglm.cn)
ChatGLM-6B uses technology similar to ChatGPT, optimized for Chinese QA and dialogue. The model is trained for about 1T tokens of Chinese and English corpus, supplemented by supervised fine-tuning, feedback bootstrap, and reinforcement learning wit human feedback. With only about 6.2 billion parameters, the model is able to generate answers that are in line with human preference.
In order to facilitate downstream developers to customize the model for their own application scenarios, we also implements an parameter-efficient tuning method based on [P-Tuning v2](https://github.com/THUDM/P-tuning-v2)[(Guidelines)](ptuning/README_en.md). Tuning requires at least 7GB of GPU memory at INT4 quantization level.
ChatGLM-6B weights are **completely open** for academic research, and **free commercial use** is also allowed after completing the [questionnaire](https://open.bigmodel.cn/mla/form).
**[2023/07/25]** Release [CodeGeeX2](https://github.com/THUDM/CodeGeeX2), which is based on ChatGLM2-6B and trained on more code data. It has the following features:
* **More Powerful Coding Capabilities**: CodeGeeX2-6B has been further pre-trained on 600B code tokens, which has been comprehensively improved in coding capability compared to the first-generation. On the [HumanEval-X](https://huggingface.co/datasets/THUDM/humaneval-x) benchmark, all six languages have been significantly improved (Python +57%, C++ +71%, Java +54%, JavaScript +83%, Go +56%, Rust +321\%), and in Python it reached 35.9% of Pass@1 one-time pass rate, surpassing the larger StarCoder-15B.
* **More Useful Features**: Inheriting the ChatGLM2-6B model features, CodeGeeX2-6B better supports both Chinese and English prompts, maximum 8192 sequence length, and the inference speed is significantly improved compared to the first-generation. After quantization, it only needs 6GB of GPU memory for inference, thus supports lightweight local deployment.
* **Comprehensive AI Coding Assistant**: The backend of CodeGeeX plugin ([VS Code](https://marketplace.visualstudio.com/items?itemName=aminer.codegeex), [Jetbrains](https://plugins.jetbrains.com/plugin/20587-codegeex)) is upgraded, supporting 100+ programming languages, and adding practical functions such as infilling and cross-file completion. Combined with the "Ask CodeGeeX" interactive AI coding assistant, it can be used to solve various programming problems via Chinese or English dialogue, including but not limited to code summarization, code translation, debugging, and comment generation, which helps increasing the efficiency of developpers.
**[2023/06/25]** Release [ChatGLM2-6B](https://github.com/THUDM/ChatGLM2-6B), the second-generation version of ChatGLM-6B. It retains the smooth conversation flow and low deployment threshold of the first-generation model, while introducing the following new features:
1.**Stronger Performance**: Based on the development experience of the first-generation ChatGLM model, we have fully upgraded the base model of ChatGLM2-6B. ChatGLM2-6B uses the hybrid objective function of [GLM](https://github.com/THUDM/GLM), and has undergone pre-training with 1.4T bilingual tokens and human preference alignment training. The [evaluation results](README.md#evaluation-results) show that, compared to the first-generation model, ChatGLM2-6B has achieved substantial improvements in performance on datasets like MMLU (+23%), CEval (+33%), GSM8K (+571%), BBH (+60%), showing strong competitiveness among models of the same size.
2.**Longer Context**: Based on [FlashAttention](https://github.com/HazyResearch/flash-attention) technique, we have extended the context length of the base model from 2K in ChatGLM-6B to 32K, and trained with a context length of 8K during the dialogue alignment, allowing for more rounds of dialogue. However, the current version of ChatGLM2-6B has limited understanding of single-round ultra-long documents, which we will focus on optimizing in future iterations.
3.**More Efficient Inference**: Based on [Multi-Query Attention](http://arxiv.org/abs/1911.02150) technique, ChatGLM2-6B has more efficient inference speed and lower GPU memory usage: under the official implementation, the inference speed has increased by 42% compared to the first generation; under INT4 quantization, the dialogue length supported by 6G GPU memory has increased from 1K to 8K.
Fore more information, please refer to [ChatGLM2-6B](https://github.com/THUDM/ChatGLM2-6B).
**[2023/05/17]** Release [VisualGLM-6B](https://github.com/THUDM/VisualGLM-6B), a multimodal conversational language model supporting image understanding.
You can run the command line and web demo through [cli_demo_vision.py](cli_demo_vision.py) and [web_demo_vision.py](web_demo_vision.py) in the repository. Note that VisualGLM-6B requires additional installation of [SwissArmyTransformer](https://github.com/THUDM/SwissArmyTransformer/) and torchvision. For more information, please refer to [VisualGLM-6B](https://github.com/THUDM/VisualGLM-6B).
**[2023/05/15]** Update the checkpoint of v1.1 version, add English instruction data for training to balance the proportion of Chinese and English data, which solves the phenomenon of Chinese words mixed in English answers .
* [ChatGLM-MNN](https://github.com/wangzhaode/ChatGLM-MNN): An MNN-based implementation of ChatGLM-6B C++ inference, which supports automatic allocation of computing tasks to GPU and CPU according to the size of GPU memory
* [JittorLLMs](https://github.com/Jittor/JittorLLMs): Running ChatGLM-6B in FP16 with a minimum of 3G GPU memory or no GPU at all, with Linux, windows, and Mac support
* [InferLLM](https://github.com/MegEngine/InferLLM): Lightweight C++ inference, which can realize real-time chat on local x86 and Arm processors, and can also run in real time on mobile phones. It only requires 4G of running memory.
* [langchain-ChatGLM](https://github.com/imClumsyPanda/langchain-ChatGLM): ChatGLM application based on langchain, realizing Q&A based on extensible knowledge base
* [Wenda](https://github.com/l15y/wenda): Large-scale language model call platform, based on ChatGLM-6B to achieve ChatPDF-like functions
* [chatgpt_academic](https://github.com/binary-husky/chatgpt_academic): An academic writing and programming toolbox that supports ChatGLM-6B. It has the characteristics of modularization and multi-thread calling LLM, and can call multiple LLMs in parallel.
* [glm-bot](https://github.com/initialencounter/glm-bot): Connect ChatGLM to Koishi to call ChatGLM on major chat platforms
Install the requirements with pip: `pip install -r requirements.txt`. `transformers` library version is recommended to be `4.27.1`, but theoretically any version no lower than `4.23.1` is acceptable.
In addition, if you need to run the quantified model on the CPU, you also need to install `gcc` and `openmp`. Most Linux distributions are installed by default. For Windows, you can check `openmp` when installing [TDM-GCC](https://jmeubank.github.io/tdm-gcc/). On Windows testing environment, the `gcc` version is `TDM-GCC 10.3.0`, and on Linux is `gcc 11.3.0`.
The implementation of the model is still in development. If you want to fix the used model implementation to ensure compatibility, you can add the `revision="v1.1.0"` parameter in the `from_pretrained` call. `v1.1.0` is the latest version number. For a complete list of versions, see [Change Log](https://huggingface.co/THUDM/chatglm-6b#change-log).
The above code will automatically download the model implementation and checkpoints by [transformers](https://github.com/huggingface/transformers). The full model implementation can be found at [Hugging Face Hub](https://huggingface.co/THUDM/chatglm-6b). If your network environment is poor, downloading model parameters may take a long time or even fail. At this point, you can download the model to the local first, and then load it from the local.
To download models from Hugging Face Hub, you need to [install Git LFS](https://docs.github.com/zh/repositories/working-with-files/managing-large-files/installing-git-large-file-storage) , then run
```Shell
git clone https://huggingface.co/THUDM/chatglm-6b
```
After downloading the model locally, replace `THUDM/chatglm-6b` in the above code with the path of your local `chatglm-6b` folder to load the model locally.
**Optional**: The implementation of the model is still in development. If you want to fix the used model implementation to ensure compatibility, you can execute
Thanks to [@AdamBear](https://github.com/AdamBear) for implementing a web demo based on Streamlit, see [#117](https://github.com/THUDM/ChatGLM-6B/pull/117 ).
The command runs an interactive program in the shell. Type your instruction in the shell and hit enter to generate the response. Type `clear` to clear the dialogue history and `stop` to terminate the program.
By default, the model parameters are loaded with FP16 precision, which require about 13GB of GPU memory. It your GPU memory is limited, you can try to load the model parameters with quantization:
After 2 to 3 rounds of dialogue, the GPU memory usage is about 10GB under 8-bit quantization, and only 6GB under 4-bit quantization. As the number of dialogue rounds increases, the corresponding GPU memory consumption also increases. Due to the use of relative position encoding, ChatGLM-6B theoretically supports an infinitely long context-length, but the performance will gradually decline after the total length exceeds 2048 (training length).
Model quantization brings a certain performance decline. After testing, ChatGLM-6B can still perform natural and smooth generation under 4-bit quantization. using [GPT-Q](https://arxiv.org/abs/2210.17323) etc. The quantization scheme can further compress the quantization accuracy/improve the model performance under the same quantization accuracy. You are welcome to submit corresponding Pull Requests.
The quantization costs about 13GB of CPU memory to load the FP16 model. If your CPU memory is limited, you can directly load the quantized model, which costs only 5.2GB CPU memory:
If your encounter the error `Could not find module 'nvcuda.dll'` or `RuntimeError: Unknown platform: darwin`(MacOS), please [load the model locally](README_en.md#load-the-model-locally).
For Macs (and MacBooks) with Apple Silicon, it is possible to use the MPS backend to run ChatGLM-6B on the GPU. First, you need to refer to Apple's [official instructions](https://developer.apple.com/metal/pytorch) to install PyTorch-Nightly. (The correct version number should be 2.1.0.dev2023xxxx, not 2.0.0).
Currently you must [load the model locally](README_en.md#load-the-model-locally) on MacOS. Change the code to load the model from your local path, and use the mps backend:
Loading a FP16 ChatGLM-6B model requires about 13GB of memory. Machines with less memory (such as a MacBook Pro with 16GB of memory) will use the virtual memory on the hard disk when there is insufficient free memory, resulting in a serious slowdown in inference speed. At this time, a quantized model such as chatglm-6b-int4 can be used. Because the quantized kernel on the GPU is written in CUDA, it cannot be used on MacOS, and can only be inferred using the CPU:
If you have multiple GPUs, but the memory size of each GPU is not sufficient to accommodate the entire model, you can split the model across multiple GPUs.
This will deploy the model onto two GPUs for inference. You can change `num_gpus` to the number of GPUs you want to use. By default, the model is split evenly, but you can also specify the `device_map` parameter to customize the splitting.
Parameter-efficient tuning based on [P-tuning v2](https://github.com/THUDM/P-tuning-v2). See [ptuning/README.md](ptuning/README.md) for details on how to use it.
This repository is licensed under the [Apache-2.0 License](LICENSE). The use of ChatGLM-6B model weights is subject to the [Model License](MODEL_LICENSE)。
## Citation
If you find our work useful, please consider citing the following papers:
```
@inproceedings{
zeng2023glm-130b,
title={{GLM}-130B: An Open Bilingual Pre-trained Model},
author={Aohan Zeng and Xiao Liu and Zhengxiao Du and Zihan Wang and Hanyu Lai and Ming Ding and Zhuoyi Yang and Yifan Xu and Wendi Zheng and Xiao Xia and Weng Lam Tam and Zixuan Ma and Yufei Xue and Jidong Zhai and Wenguang Chen and Zhiyuan Liu and Peng Zhang and Yuxiao Dong and Jie Tang},
booktitle={The Eleventh International Conference on Learning Representations (ICLR)},
