InfoQuant: Shaping Activation Distributions for Low-Bit LLM Quantization

Name: InfoQuant: Shaping Activation Distributions for Low-Bit LLM Quantization
Uploaded: 2026-10-23T09:57:59.341914+00:00
Duration: 1211 s
Description: INFOQUANT is a train-free method to improve low-bit quantization for large language models by reshaping activation distributions. It uses three components—PSOT, ASOT, and LAC—to transform activations into distributions with smaller numerical ranges and higher dispersion. This approach significantly reduces quantization error, leading to better performance and hardware efficiency.

INFOQUANT is a train-free method to improve low-bit quantization for large language models by reshaping activation distributions. It uses three components—PSOT, ASOT, and LAC—to transform activations into distributions with smaller numerical ranges and higher dispersion. This approach significantly reduces quantization error, leading to better performance and hardware efficiency.

Open the full episode