Xinwei Niu

Xinwei Niu

Research Assistant at Encode Lab

SparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference

Available as arXiv preprint, arXiv:2610.12327, 2026

Qitong Wang, Xinwei Niu, Mingluo Su, Shanwei Zhao, Shiai Zhu, and Huan Wang.

arXiv preprint · October 8, 2026

Abstract

The memory-bound nature of the decoding stage of large language model (LLM) inference incurs significant latency. Layer-wise training-free network pruning approaches guided by the Hessian have been a prominent solution to this problem, as pruning reduces the number of nonzero parameters read from memory during decoding.

Nevertheless, typical methods in this line compute the Hessian using pre-collected natural sequences, whereas the model is fed self-generated tokens during decoding, creating a distribution shift between the two sequences. The Hessian calculated on the natural sequence is different from that calculated on the generated sequence. We observe that this discrepancy causes the activation distribution during generation to deviate from that used for pruning, further hurting the pruned model performance.

Moreover, most existing LLM pruning methods that bring actual speedup primarily target the sparse matrix-matrix (SpMM) multiplication, providing limited support for the sparse matrix-vector (SpMV) operations, which dominate decoding. To solve these problems, we introduce SparseDecoding, a principled decoding-aware pruning framework tailored for accurate and efficient LLM decoding.

Specifically, at the algorithmic axis, SparseDecoding constructs calibration matrices from layer-wise activations collected during the dense-model autoregressive generation, excluding prefill, thereby aligning the pruning objective with the decoding activations. At the system axis, we develop an optimized N:M sparse matrix-vector kernel with bitmask indexing and fixed-step traversal.

Substantial empirical results on representative LLMs (Llama-3.1-8B, Llama-3.3-70B, Qwen3-14B / 32B) demonstrate that our method consistently outperforms standard fixed-text calibration on the long-form generation benchmarks while achieving up to 1.48x end-to-end wall-clock decoding speedup on A100 GPUs.

arXiv · PDF

Figures

Relative activation discrepancy across decoding steps, with an inset showing the first 40 steps. (a) Activation discrepancy during decoding Sparse prefill speedup is above the dense baseline, while sparse decoding speedup is below it across sequence lengths. (b) Prefill and decoding speedup
Figure 1. Motivation for decoding-aware sparse inference: activation discrepancies emerge early in generation, and sparse kernels designed for prefill do not necessarily accelerate decoding.
Comparison of standard N:M pruning and SparseDecoding, from calibration data and mask selection to decode-time execution.
Figure 2. SparseDecoding uses generated-token activations for calibration, excludes prefill statistics, and executes pruned weights with an optimized N:M sparse matrix-vector kernel.

Figures from SparseDecoding on arXiv. Select a figure to view it at full size.

Recommended citation: Qitong Wang, Xinwei Niu, Mingluo Su, Shanwei Zhao, Shiai Zhu, and Huan Wang. (2026). SparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference. arXiv preprint arXiv:2610.12327.
Download Paper