Diagonal-Tiled Mixed-Precision Attention for Efficient Low-Bit MXFP Inference
This work addresses the problem of efficient low-bit inference for large language models, offering a domain-specific optimization that is incremental in nature.
The paper tackles the high inference cost of transformer-based large language models by introducing a low-bit mixed-precision attention kernel using the MXFP data format, achieving significant speedup on NVIDIA B200 GPUs with negligible performance degradation.
Transformer-based large language models (LLMs) have demonstrated remarkable performance across a wide range of real-world tasks, but their inference cost remains prohibitively high due to the quadratic complexity of attention and the memory bandwidth limitations of high-precision operations. In this work, we present a low-bit mixed-precision attention kernel using the microscaling floating-point (MXFP) data format, utilizing the computing capability on next-generation GPU architectures. Our Diagonal-Tiled Mixed-Precision Attention (DMA) incorporates two kinds of low-bit computation at the tiling-level, and is a delicate fused kernel implemented using Triton, exploiting hardware-level parallelism and memory efficiency to enable fast and efficient inference without compromising model performance. Extensive empirical evaluations on NVIDIA B200 GPUs show that our kernel maintains generation quality with negligible degradation, and meanwhile achieves significant speedup by kernel fusion. We release our code at https://github.com/yifu-ding/MP-Sparse-Attn.