Inference optimizations are critical for improving user experience and reducing infrastructure costs and power consumption. In this article, we illustrate a form of dynamic execution known as speculative sampling to reduce the overall latency of text generation and compare it with standard autoregressive sampling. This can be used together with model-based optimizations (e.g. quantization) to provide an optimized solution. Both sampling methods make use of KV caching. A Jupyter notebook and some sample executions are provided.
@article{arxiv.2311.04951,
title = {Leveraging Speculative Sampling and KV-Cache Optimizations Together for Generative AI using OpenVINO},
author = {Haim Barad and Ekaterina Aidova and Yury Gorbachev},
journal= {arXiv preprint arXiv:2311.04951},
year = {2024}
}
Comments
Code available at https://github.com/openvinotoolkit/openvino_notebooks/tree/latest/notebooks/speculative-sampling