GLM is a General Language Model pretrained with an autoregressive blank-filling objective and can be finetuned on various natural language understanding and generation tasks.

Please refer to our paper for a detailed description of GLM:

GLM: General Language Model Pretraining with Autoregressive Blank Infilling (ACL 2022)

Zhengxiao Du*, Yujie Qian*, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, Jie Tang (*: equal contribution)

News: We release ChatGLM-6B, an open pre-trained language model with 6 billion parameters optimized for Chinese QA and dialogue based on the GLM framework.

You can download the pretrained models used in the paper from OneDrive or Tsinghua-Cloud.

Unzip the downloaded file into a local folder and set CHECKPOINT_PATH in the corresponding scripts to the folder path.

dev set, single model, single-task finetuning

CNN/Daily Mail (test set, no additional data used)

XSum (test set, no additional data used)

test set, zero-shot

You can access GLM models via HuggingFace Hub. Please install transformers>=4.23.1 and find all the available models here.

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tokenizer = AutoTokenizer.from_pretrained("THUDM/glm-10b", trust_remote_code=True)
model = AutoModelForSeq2SeqLM.from_pretrained("THUDM/glm-10b", trust_remote_code=True)
model = model.half().cuda()
model.eval()

# Inference
inputs = tokenizer("Ng is an adjunct professor at [MASK] (formerly associate professor and Director of its Stanford AI Lab or SAIL ). Also a pioneer in online education, Ng co-founded Coursera and deeplearning.ai.", return_tensors="pt")
inputs = tokenizer.build_inputs_for_generation(inputs, max_gen_length=512)
inputs = inputs.to('cuda')
outputs = model.generate(**inputs, max_length=512, eos_token_id=tokenizer.eop_token_id)
print(tokenizer.decode(outputs[0].tolist()))

# Training
inputs = tokenizer(
    ["Tsinghua University is located in [MASK].", "One minus one equals zero, is it correct? Answer: [MASK]"],
    return_tensors="pt", padding=True)
inputs = tokenizer.build_inputs_for_generation(inputs, targets=["Beijing", "No"], max_gen_length=8, padding=False)
inputs = inputs.to('cuda')
outputs = model(**inputs)
loss = outputs.loss
logits = outputs.logits

from transformers import AutoTokenizer, AutoModelForMultipleChoice
tokenizer = AutoTokenizer.from_pretrained("THUDM/glm-10b", trust_remote_code=True)
model = AutoModelForMultipleChoice.from_pretrained("THUDM/glm-10b", trust_remote_code=True)
model = model.half().cuda()
model.eval()

inputs = tokenizer(["Tsinghua University is located in [MASK].",
                    "One minus one equals zero, is it correct? Answer: [MASK]"], return_tensors="pt", padding=True)
choices = [["Beijing", "Shanghai"], ["Yes", "No"]]
inputs = tokenizer.build_inputs_for_multiple_choice(inputs, choices)
inputs = inputs.to('cuda')
outputs = model(**inputs)
logits = outputs.logits

You can also convert the finetuned checkpoints with scripts/convert_glm_checkpoint_to_transformers.py.

We prepare two docker images based on CUDA 10.2 and CUDA 11.2. You can pull the pre-built images from Docker Hub and run with docker v19.03

docker run --gpus all --rm -it --ipc=host zxdu20/glm-cuda102

or replace glm-cuda102 with glm-cuda112.

You can also modify the image according to your requirements in docker/cuda102.dockerfile and build the image yourself

  docker build -f cuda102.dockerfile . -t glm-cuda102

Please first install PyTorch (we use 1.7.0) and apex, and then install other dependencies by pip install -r requirements.txt

git clone https://github.com/THUDM/GLM
cd GLM

If your encounter the CUDA out of memory error, which means you GPU memory is limited, you can try the model parallelism to divide the parameters into multiple GPUs. Take the two-way model parallelism as an example. First run change_mp.py to divide the checkpoint:

python change_mp.py path_to_the_checkpoint 2

Then update the checkpoint path in the model config file (such as config_tasks/model_blocklm_10B.sh) and change MP_SIZE in the script (such as scripts/ds_finetune_superglue.sh) to 2.

We provide scripts for finetuning GLM on some downstream tasks.

• Change CHECKPOINT_PATH to your local path. Run the following script

bash scripts/generate_block.sh \
     config_tasks/model_blocklm_10B_chinese.sh

Some models (GLM-2B, GLM-10B, and GLM-10B-Chinese) use three different mask tokens: [MASK] for short blank filling, [sMASK] for sentence filling, and [gMASK] for left-to-right generation.

You can also add multiple [MASK] and [sMASK] in a single example. The model will fill the blanks one by one from left to right. The answer to each blank always begins with a special <|startofpiece|>.

• Download the SuperGlue data and check the experiment setup in scripts/ds_finetune_superglue.sh. Note that DATA_ROOT, CHECKPOINT_PATH, SAVE_PATH need to be changed to your local path. You may also change the batch-size and nproc_per_node according to your available hardware.

• Run the following script (use the COPA dataset as an example)

bash scripts/ds_finetune_superglue.sh \
     config_tasks/model_blocklm_10B.sh \
     config_tasks/task_copa.sh

• We also implement P-Tuning in our code. Run the following script to integrate p-tuning:

bash scripts/ds_finetune_superglue_prompt.sh \
     config_tasks/model_blocklm_10B.sh \
     config_tasks/task_copa.sh

• To apply GLM to a new NLU dataset with cloze-filling finetuning, implement a DataProcessor in tasks/superglue/dataset.py for data loading and add a PVP in tasks/superglue/pvp.py for the cloze question. More details can be found here.

• Download the Gigaword , CNN/Daily Mail or XSum dataset and check the experiment setup in scripts/ds_finetune_seq2seq.sh. Change DATA_ROOT, CHECKPOINT_PATH, SAVE_PATH to your local path.

• Run the following script (use the CNN/Daily Mail dataset as an example)

  bash scripts/ds_finetune_seq2seq.sh \ 
     config_tasks/model_blocklm_10B.sh \ 
     config_tasks/seq_cnndm_org.sh
  

• The summaries are written into ./runs/experiment_name/test.jsonl.hyps. The references are written into test.jsonl.refs in the same directory. For calculating rouge, install file2rouge and download Stanford CoreNLP from here. Run the following script

  bash scripts/evaluate_seq2seq.sh \
   ./runs/experiment_name/test.jsonl.hyps ./runs/experiment_name/test.jsonl.refs
  

Process your seq2seq data into {split}.source and {split}.target, with each line being the context or the target of a sample, and split being train, val, and test.

Run the following script

bash scripts/ds_finetune_seq2seq.sh \ 
   config_tasks/model_blocklm_10B.sh \ 
   config_tasks/seq_customization.sh

You can specify the hyperparameters in config_tasks/seq_customization.sh and config_tasks/config_blocklm_10B_cnndm.json

bash scripts/evaluate_multichoice.sh config_tasks/model_blocklm_10B.sh

Note that CHECKPOINT_PATH and DATA_PATH need to be changed to your local path.

The format of each line of the data file should be

{"inputs_pretokenized": "Context and question here", "choices_pretokenized": ["Choice 1", "Choice 2", "Choice 3"], "label": int}

• Download the LAMBADA data and change DATA_ROOT, CHECKPOINT_PATH in scripts/evaluate_lm.sh
• Run the following script

bash scripts/evaluate_lm.sh \ 
     config_tasks/model_blocklm_large_generation.sh \
     config_tasks/zero_lambada.sh 

• Download our test set of wikibook or Wikitext103 dataset and change DATA_ROOT, CHECKPOINT_PATH in scripts/evaluate_lm.sh
• Run the following script

  bash scripts/evaluate_lm.sh \ 
     config_tasks/model_blocklm_large_generation.sh \
     config_tasks/zero_wikitext.sh 

• Download the Yahoo dataset and check the experiment setup in scripts/finetune_blank.sh. Change DATA_ROOT, CHECKPOINT_PATH, SAVE_PATH to your local path.

• Run the following script

bash scripts/finetune_blank.sh \ 
     config_tasks/model_blocklm_large.sh \ 
     config_tasks/seq_blank.sh

Run the following script to pre-train the GLM-Large model

bash scripts/ds_pretrain_nvidia.sh config/ds_block_large.sh

The script scripts/ds_pretrain_nvidia.sh launches the training program with DeepSpeed. You should change NUM_WORKERS and NUM_GPUS_PER_WORKER to the number of workers and the number of gpus per worker. Also change HOST_FILE_PATH to the path to an OpenMPI-style hostfile. More details about DeepSpeed launcher can be found here.

The file config/ds_block_large.sh defines the hyperparameters for pretraining. Most of the arguments are fairly self-explanatory. Specifically, --train-data can be multiple keywords defined in NAMED_CORPORA in data_utils/corpora.py. The hyperparameters of the optimizer are defined in the corresponding json file under config. The semantics of the json file can be found here.

Part of the code is based on Megatron-LM and PET.

Please cite our paper if you find this code useful for your research:

@article{DBLP:conf/acl/DuQLDQY022,
  author    = {Zhengxiao Du and
               Yujie Qian and
               Xiao Liu and
               Ming Ding and
               Jiezhong Qiu and
               Zhilin Yang and
               Jie Tang},
  title     = {{GLM:} General Language Model Pretraining with Autoregressive Blank Infilling},
  booktitle = {Proceedings of the 60th Annual Meeting of the Association for Computational
               Linguistics (Volume 1: Long Papers), {ACL} 2022, Dublin, Ireland,
               May 22-27, 2022},
  pages     = {320--335},
  publisher = {Association for Computational Linguistics},
  year      = {2022},
}