Zero-Shot Dynamic Quantization for Transformer Inference

Code

• github.com/2023-MindSpore-1/ms-code-211/tree/main/DynamicQuant
  mindspore★ 0
• github.com/MindSpore-MS-Code2/code0/tree/main/DynamicQuant
  mindspore★ 0
• github.com/2023-MindSpore-4/Code3/tree/main/DynamicQuant
  mindspore★ 0
• github.com/MindSpore-paper-code-3/code8/tree/main/DynamicQuant
  mindspore★ 0

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.

Tasks

Reproductions