New Method Optimizes LLM Quantization for Performance and Quality.

Researchers have developed a new technique for quantizing large language models (LLMs) that intelligently assigns different bit-widths to individual layers. Unlike uniform quantization methods, this approach focuses on maximizing performance, specifically reducing latency, while carefully controlling the degradation of generation quality. The method leverages a layer sensitivity profile and applies it within TensorRT-LLM, customizing precision for components like Feed-Forward Networks (FFN), Attention, and the lm_head. Experiments on Gemma-3-1B demonstrated that this selective quantization can yield substantial latency reductions. For instance, optimizing FFN and lm_head layers resulted in an 11.0% latency decrease with negligible quality impact. Further adjustments allowed for speedups up to 19.1% with acceptable quality loss. The study also highlighted that while integer arithmetic benefits FFN and lm_head, it can slow down Attention for short contexts due to quantization/dequantization overhead.