Leveraging Speculative Sampling and KV-Cache Optimizations Together for Generative AI using OpenVINO
This work addresses inference optimization for generative AI users to enhance user experience and reduce costs, but it appears incremental as it combines existing techniques without claiming major breakthroughs.
The paper tackled the problem of reducing latency in text generation for generative AI by combining speculative sampling with KV-cache optimizations, achieving unspecified improvements in inference speed and efficiency.
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.