Deberta Xxlarge Fixed

🚀 DeBERTa (1.5B) fixed version

This is an updated version of deberta-v2-xxlarge that implements the AutoModelForCausalLM class, enabling text generation. The implementation is based on our paper "BERTs are Generative In-Context Learners".

🚀 Quick Start

This repository also fixes three bugs in the original HF implementation of DeBERTa:

1. We fixed the incorrect name of the output embedding weights in the checkpoint file;
2. We fixed the implementation of the enhanced mask decoder (EMD), based on the original GitHub repository;
3. We clamp the positional embeddings so that they work with long sequence lengths.

✨ Features

• Updated to implement the AutoModelForCausalLM class for text generation.
• Fixed three bugs in the original HF implementation of DeBERTa.

💻 Usage Examples

Basic Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("ltg/deberta-xxlarge-fixed", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("ltg/deberta-xxlarge-fixed", trust_remote_code=True).cuda().eval()

prompt = """German: Hallo, wie geht es Ihnen heute?
English:"""
prompt = prompt.replace('\n', '\\n ')
input_ids = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).input_ids.cuda()

prediction = model.generate(
    input_ids,
    num_beams=4,
    do_sample=False,
    use_cache=None,
    max_new_tokens=64,
    eos_token_id=tokenizer(".\\", add_special_tokens=False).input_ids[1:]
)
prediction = prediction[0, input_ids.size(1):]
prediction = tokenizer.decode(prediction).rstrip('\\')

# Expected output: "Hello, how are you doing today?"
print(prediction)

📄 License

This project is licensed under the MIT license.

📚 Documentation

Citation

If you find DeBERTa useful for your work, please cite the following paper:

@inproceedings{
  samuel2024berts,
  title={{BERT}s are Generative In-Context Learners},
  author={David Samuel},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
  year={2024},
  url={https://openreview.net/forum?id=BCA9NMZkLS}
}

@inproceedings{he2021deberta,
  title={{DeBERTa}: Decoding-enhanced {BERT} with disentangled attention},
  author={Pengcheng He and Xiaodong Liu and Jianfeng Gao and Weizhu Chen},
  booktitle={International Conference on Learning Representations},
  year={2021},
  url={https://openreview.net/forum?id=XPZIaotutsD}
}