ADAngel: Accelerating Arbitrary-Precision Quantized LLMs with Adaptive Computing Mapping

Arbitrary-Precision Quantization (APQ), which uses asymmetric bit-widths for weights and activations (e.g., W4A8), is a prevalent technique for LLM inference because of its excellent accuracy-performance balance. APQ transforms the general matrix multiplications (GEMM), the core of LLM computation, into mixed-precision GEMM (mpGEMM) whose two operand matrices have different quantization bit-widths. However, we identify that the computation paradigms of mpGEMM in current APQ LLM inference systems are sub-optimal because the shapes and bit-widths of mpGEMM tasks in APQ LLM are highly variable, whereas existing static and workload-unaware paradigms can only accelerate mpGEMM tasks with the same or similar shapes and bit-widths. Based on this finding, we propose ADAngel, a framework for creating a workload-adaptive mpGEMM computation core for target LLMs. The theoretical foundation of ADAngel is the DPR (Decomposition-Partial Product-Reconstruction) computation model, which enables systematic generation of a diverse portfolio of mpGEMM algorithms by specifying different bit-partition schemes. Guided by this model, ADAngel constructs a Computation Strategy Set comprising several highly optimized mpGEMM kernels, and exhaustively analyzes the strategy set to create an Oracle Policy Map, which enables a lightweight dispatcher to select and execute the optimal kernel for runtime mpGEMM tasks with negligible overhead. Our evaluation shows that the ADAngel-specialized engine achieves up to a 5.10× speedup in decode throughput over llama.cpp; while in the prefill stage, it demonstrates its adaptivity by delivering speedups ranging from 1.17× to 2.38× over TensorRT-LLM in Time-To-First-Token (TTFT).