PhaseQuant: Mitigating Activation Quantization Collapse via Residual-Stream Dynamics

DynamicPTQ introduces a phase-aware mixed-precision activation strategy for low-bit LLM inference. It uses residual stream dynamics metrics, Jump Ratio and Historical Feature SNR, to identify layers most sensitive to quantization loss. By assigning 8-bit precision only to these sensitive layers while keeping others at 4-bit, the method consistently improves performance across various models and datasets while maintaining efficiency.

Open the full episode

More from AI Radio