QUICK: Quantization-aware Interleaving and Conflict-free Kernel for efficient LLM inference
This is an incremental improvement for efficient LLM inference on GPUs, addressing a specific bottleneck in existing quantization kernels.
The paper tackles the shared memory bank-conflict problem in mixed precision matrix multiplication kernels for quantized LLM inference by introducing QUICK, a group of optimized CUDA kernels that interleave quantized weight matrices offline to skip shared memory write-back after dequantization, achieving up to 1.91x speedup over existing kernels on larger batches and up to 1.94x throughput gain on representative LLM models across NVIDIA GPUs.
We introduce QUICK, a group of novel optimized CUDA kernels for the efficient inference of quantized Large Language Models (LLMs). QUICK addresses the shared memory bank-conflict problem of state-of-the-art mixed precision matrix multiplication kernels. Our method interleaves the quantized weight matrices of LLMs offline to skip the shared memory write-back after the dequantization. We demonstrate up to 1.91x speedup over existing kernels of AutoAWQ on larger batches and up to 1.94x throughput gain on representative LLM models on various NVIDIA GPU devices.