KLLM: Fast LLM Inference with K-Means Quantization (2507.23035v1)

Abstract: LLM inference poses significant challenges due to its intensive memory and computation demands. Weight and activation quantization (WAQ) offers a promising solution by reducing both memory footprint and arithmetic complexity. However, two key challenges remain in the existing WAQ designs. (1) Traditional WAQ designs rely on uniform integer-based quantization for hardware efficiency, but this often results in significant accuracy degradation at low precision. K-Means-based quantization, a non-uniform quantization technique, achieves higher accuracy by matching the Gaussian-like distributions of weights and activations in LLMs. However, its non-uniform nature prevents direct execution on low-precision compute units, requiring dequantization and floating-point matrix multiplications (MatMuls) during inference. (2) Activation outliers further hinder effective low-precision WAQ. Offline thresholding methods for outlier detection can lead to significant model performance degradation, while existing online detection techniques introduce substantial runtime overhead. To address the aforementioned challenges and fully unleash the potential of WAQ with K-Means quantization for LLM inference, in this paper, we propose KLLM, a hardware-software co-design framework. KLLM features an index-based computation scheme for efficient execution of MatMuls and nonlinear operations on K-Means-quantized data, which avoids most of the dequantization and full-precision computations. Moreover, KLLM incorporates a novel outlier detection engine, Orizuru, that efficiently identifies the top-$k$ largest and smallest elements in the activation data stream during online inference. Extensive experiments show that, on average, KLLM achieves speedups of 9.67x, 7.03x and energy efficiency improvements of 229.50x, 150.21x compared to the A100 GPU and Atom, respectively.