Abstract
We introduce a novel run-time method for significantly reducing the accuracy loss associated with quantizing BERT-like models to 8-bit integers. Existing methods for quantizing models either modify the training procedure, or they require an additional calibration step to adjust parameters that also requires a selected held-out dataset. Our method permits taking advantage of quantization without the need for these adjustments. We present results on several NLP tasks demonstrating the usefulness of this technique.- Anthology ID:
- 2022.emnlp-industry.45
- Volume:
- Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track
- Month:
- December
- Year:
- 2022
- Address:
- Abu Dhabi, UAE
- Editors:
- Yunyao Li, Angeliki Lazaridou
- Venue:
- EMNLP
- SIG:
- Publisher:
- Association for Computational Linguistics
- Note:
- Pages:
- 451–457
- Language:
- URL:
- https://aclanthology.org/2022.emnlp-industry.45
- DOI:
- 10.18653/v1/2022.emnlp-industry.45
- Cite (ACL):
- Yousef El-kurdi, Jerry Quinn, and Avi Sil. 2022. Zero-Shot Dynamic Quantization for Transformer Inference. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 451–457, Abu Dhabi, UAE. Association for Computational Linguistics.
- Cite (Informal):
- Zero-Shot Dynamic Quantization for Transformer Inference (El-kurdi et al., EMNLP 2022)
- PDF:
- https://preview.aclanthology.org/emnlp-22-attachments/2022.emnlp-industry.45.pdf