🌐 Blog • ⏬ Download Model • 🪧 Demo • ✉️ Email • 📃 Paper [ICLR 2023]

💬 Google Group (Updates) or Wechat Group or Slack channel (Discussions)

GLM-130B: An Open Bilingual Pre-Trained Model

GLM-130B is an open bilingual (English & Chinese) bidirectional dense model with 130 billion parameters, pre-trained using the algorithm of General Language Model (GLM). It is designed to support inference tasks with the 130B parameters on a single A100 (40G * 8) or V100 (32G * 8) server. With INT4 quantization, the hardware requirements can further be reduced to a single server with 4 * RTX 3090 (24G) with almost no performance degradation. As of July 3rd, 2022, GLM-130B has been trained on over 400 billion text tokens (200B each for Chinese and English) and it has the following unique features:

• Bilingual: supports both English and Chinese.
• Performance (EN): better than GPT-3 175B (+4.0%), OPT-175B (+5.5%), and BLOOM-176B (+13.0%) on LAMBADA and slightly better than GPT-3 175B (+0.9%) on MMLU.
• Performance (CN): significantly better than ERNIE TITAN 3.0 260B on 7 zero-shot CLUE datasets (+24.26%) and 5 zero-shot FewCLUE datasets (+12.75%).
• Fast Inference: supports fast inference on both SAT and FasterTransformer (up to 2.5X faster) with a single A100 server.
• Reproducibility: all results (30+ tasks) can be easily reproduced with open-sourced code and model checkpoints.
• Cross-Platform: supports training and inference on NVIDIA, Hygon DCU, Ascend 910, and Sunway (Will be released soon).

This repository mainly focus on the evaluation of GLM-130B. If you find our work and our open-sourced efforts useful, ⭐️ to encourage our following development! :)

News

• [2023.03.14] We are happy to introduce ChatGLM, a bilingual dialogue language model based on GLM-130B, and its open-sourced version ChatGLM-6B which can be run under only 6GB GPU memory!
• [2023.01.21] GLM-130B has been accepted to ICLR 2023!
• [2022.10.06] Our paper for GLM-130B is out!
• [2022.08.24] We are proud to publish the quantized version for GLM-130B. While preserving the activation precision as FP16, the model weights can be quantized to as low as INT4 with almost no degradation of performance, further reducing the hardware requirements of the GLM-130B to a single server with 4 * RTX 3090 (24G)! See Quantization of GLM-130B for details.

For smaller models, please find monolingual GLMs (English: 10B/2B/515M/410M/335M/110M, Chinese: 10B/335M) and an 1B multilingual GLM (104 languages).

Getting Started

Environment Setup

Hardware

It is recommended to use the an A100 (40G * 8) server, as all GLM-130B evaluation results (~30 tasks) reported can be easily reproduced with a single A100 server in about half a day. With INT8/INT4 quantization, efficient inference on a single server with 4 * RTX 3090 (24G) is possible, see Quantization of GLM-130B for details. Combining quantization and weight offloading techniques, GLM-130B can also be inferenced on servers with even smaller GPU memory, see Low-Resource Inference for details.

Software

The GLM-130B code is built on the top of SAT. We recommend using Miniconda to manage your environment and installing additional dependencies via pip install -r requirements.txt. Here are the recommended environment configurations:

• Python 3.9+ / CUDA 11+ / PyTorch 1.10+ / DeepSpeed 0.6+ / Apex (installation with CUDA and C++ extensions is required, see here)
• SwissArmyTransformer>=0.2.11 is required for quantization

Model weights

Download the GLM-130B’s model checkpoint from here, make sure all 60 chunks are downloaded completely, then use the following command to merge them into a single archive file and extract it:

cat glm-130b-sat.tar.part_* > glm-130b-sat.tar
tar xvf glm-130b-sat.tar

Set CHECKPOINT_PATH in configs/model_glm_130b.sh to the path of the extracted folder. Since the checkpoint file is up to 260G, it is recommended to use the SSD or RAM disk to reduce the checkpoint loading time. Since the checkpoint we distribute is in 8-way tensor parallel, a conversion scripts is also provided if you need to change the tensor parallel dimension.

python tools/convert_tp.py \
    --input-folder <SRC_CKPT_PATH>  \
    --output-folder <DST_CKPT_PATH> \
    --target-tp <TARGET_TP>

Left-To-Right Generation / Blank Filling

bash scripts/generate.sh --input-source interactive

You can also specify an input file by --input-source input.txt.

GLM-130B uses two different mask tokens: [MASK] for short blank filling and [gMASK] for left-to-right long text generation. When the input does not contain any MASK token, [gMASK] will be automatically appended to the end of the text.

Evaluation

We use the YAML file to define tasks. Specifically, you can add multiple tasks or folders at a time for evaluation, and the evaluation script will automatically collect all YAML files under those folders recursively.

bash scripts/evaluate.sh task1.yaml task2.yaml dir1 dir2 ...

Download our evaluation dataset here, and set DATA_PATH in scripts/evaluate.sh to your local dataset directory. The task folder contains the YAML files for 30+ tasks we evaluated for GLM-130B. Take the CoLA task for example, run bash scripts/evaluate.sh tasks/bloom/glue_cola.yaml, which outputs an accuracy of ~65% for the best prompt and ~57% for the median.

MultiChoiceTaskConfig(name='glue_cola', type=<TaskType.MULTICHOICE: 'mul'>, path='/thudm/LargeScale/data/zeroshot/bloom/glue_cola', module=None, metrics=['Accuracy'], use_task_mask=False, use_multitask_encoding=False, unidirectional=False, max_seq_length=2048, file_pattern={'validation': '**/validation.jsonl'}, micro_batch_size=8)
Evaluating task glue_cola:
  Evaluating group validation:
      Finish Following_sentence_acceptable/mul/validation.jsonl, Accuracy = 42.665
      Finish Make_sense_yes_no/mul/validation.jsonl, Accuracy = 56.951
      Finish Previous_sentence_acceptable/mul/validation.jsonl, Accuracy = 65.197
      Finish editing/mul/validation.jsonl, Accuracy = 57.622
      Finish is_this_correct/mul/validation.jsonl, Accuracy = 65.197
Evaluation results of task glue_cola:
  Group validation Accuracy: max = 65.197, median = 57.622, average = 57.526
Finish task glue_cola in 101.2s. 

Multi-node evaluation can be configured by setting HOST_FILE_PATH(required by the DeepSpeed lanucher) in scripts/evaluate_multiple_node.sh. Set DATA_PATH in scripts/evaluate_multiple_node.sh and run the following command to evaluate all the tasks in ./task directory.

bash scripts/evaluate_multiple_node.sh ./tasks

See Evaluate Your Own Tasks for details on how to add new tasks.

2.5X faster Inference using FasterTransformer

By adapting the GLM-130B model to FasterTransfomer, a highly optimized transformer model library by NVIDIA, we can reach up to 2.5X speedup on generation, see Inference with FasterTransformer for details.

License

This repository is licensed under the Apache-2.0 license. The use of GLM-130B model weights is subject to the Model License.

Citation

If you find our work useful, please consider citing GLM-130B:

@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}
}

You may also consider GLM's original work in your reference:

@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}
}