Inference Optimization

Papers on speeding up inference, decoding strategies, and serving efficiency.

Overview

This section contains 4 papers covering:

• Classifier-Free Guidance - Inference-time method for better prompt adherence
• Staged Speculative Decoding - Tree-based speculation for on-device inference
• FlashFFTConv - Optimized FFT for long-sequence convolutions
• Flash Attention - Hardware-aware attention optimization