ReQuant: Fixed-Grid Discrete Refinement for Post-Training Quantization
The episode discusses ReQuant, a post-training quantization refinement method that improves already-quantized AI models by nudging weights to neighboring grid points without retraining. Hosts highlight its backpropagation-free approach, monotonic error reduction, and significant accuracy gains on models like Llama-3 8B, making it a practical final stage for deployment.