SharQ: Bridging Activation Sparsity and FP4 Quantization for LLM Inference

2026-06-25

AI Summary

This work addresses the challenge of jointly applying low-bit floating-point (FP4) quantization and N:M semi-structured sparsity in large language models, where outlier interference disrupts block-wise scaling and naive sparsification discards medium-magnitude values, leading to coupled quantization and sparsity errors. To overcome this, the authors propose SharQ, a training-free inference method that synergistically combines online sparse-dense decomposition: it adaptively generates an N:M mask to extract an outlier-dominated sparse backbone for FP4 quantization, while constructing a dense residual relative to this backbone. Both components are computed via two FP4 GEMMs sharing weights but employing path-specific scaling factors. SharQ is the first approach to effectively integrate activation sparsity with FP4 quantization without error accumulation, supports multiple FP4 formats, and requires no calibration. Experiments show SharQ recovers 43%–63% of the accuracy gap between NVFP4 and FP16 across several large models, achieves 2.2–2.4× speedup over FP16 on RTX 5090, outperforms FP8 by 1.2–1.4× in throughput, and accelerates video generation tasks by up to 1.58×.

Abstract

Low-bit floating-point formats and semi-structured sparsity are increasingly supported by modern accelerators, yet combining them for LLM activation compression remains challenging: activations contain input-dependent outliers that dominate block scales in FP4 quantization, and directly applying N:M sparsity masks discards moderate values, coupling sparsification loss with quantization error. We introduce SharQ, a training-free inference method that bridges activation sparsity and FP4 quantization through an online sparse--dense decomposition. For each activation tensor, SharQ generates an input-adaptive N:M mask to extract an outlier-dominated sparse backbone, quantizes it to FP4, and defines a dense residual relative to the quantized sparse backbone rather than the unquantized sparse values. A sparse FP4 GEMM processes the backbone while a dense FP4 GEMM compensates for both mask-induced activation loss and sparse-path quantization error. The two paths share a single FP4 weight payload with path-specific scale views, and a fused preparation kernel absorbs mask generation, residual construction, and layer normalization into one operator. SharQ requires no calibration data, retraining, or model-specific tuning. Evaluated on Llama-3.1-8B, Qwen2.5-7B, Qwen3-30B-A3B, and Qwen3-VL-8B, SharQ recovers 43--63% of the NVFP4-to-FP16 accuracy gap across language and vision-language tasks, and generalizes across NVFP4, HiF4, and MXFP4 formats. On an RTX 5090, SharQ delivers 2.2--2.4$\times$ latency reduction over FP16 and 1.2--1.4$\times$ throughput improvement over FP8 in language model serving, and up to 1.58$\times$ speedup on Wan2.2-T2V-A14B video generation when combined with SageAttention. Our code is available at https://github.com/actypedef/SharQ.

Problem

• activation sparsity
• FP4 quantization
• LLM inference
• outlier handling
• N:M sparsity

Innovation

• activation sparsity
• FP4 quantization
• sparse-dense decomposition
• training-free inference
• low-bit LLM acceleration