Unified Static–Dynamic Pruning
for Efficient LLM Inference

VLDB 2026

1Seoul National University
* Corresponding author
Output width scaling in dense GEMM
(a) Output width scaling in dense GEMM
Sparsity scaling in SpMV
(b) Sparsity scaling in SpMV

Roofline analysis on an A10G GPU. Decode’s narrow GEMM (small output width) is memory-bound, while higher sparsity lifts SpMV compute intensity — motivating SPDP’s unified static–dynamic pruning.

Abstract

The increasing deployment of large language models (LLMs) has magnified the computational and memory bottlenecks of autoregressive decoding, where low compute intensity and bandwidth-bound kernels dominate inference cost. Weight pruning offers a promising remedy, but existing methods remain confined to either static pruning (SP)—which permanently removes redundant weights but lacks adaptivity—or dynamic pruning (DP)—which adapts to input sparsity but introduces runtime irregularity. This paper presents SPDP, a unified sparse-inference framework that integrates unstructured SP with input-adaptive DP for efficient LLM inference on GPUs.

SPDP co-designs a new Tiled-Column-wise Bitmap Compressed (Tiled-CBC) format and two complementary GPU kernels: (1) a CUDA-core spMspV kernel featuring Hybrid Activation-aware Dynamic Shared-Memory Bitmap Decoding (HAD-SMBD) for fine-grained, runtime activation skipping, and (2) a Tensor-Core SpMM kernel optimized for prefill computation. This joint format–kernel design harmonizes static and dynamic sparsity, maintaining bandwidth-efficient memory access and high compute intensity under both phases of LLM inference. Comprehensive evaluations on inference-optimized GPUs demonstrate that SPDP achieves 1.24×–1.37× average speedup (up to 2.51×) over state-of-the-art sparse frameworks such as SpInfer, while matching perplexity with up to 25% higher sparsity. SPDP advances the inference efficiency–quality Pareto frontier, showing that unified static–dynamic pruning can deliver substantial throughput and performance-per-watt improvements in large-scale LLM serving.

Motivation

Static pruning (SP) permanently removes redundant weights but is input-agnostic, so quality erodes as sparsity grows. Dynamic pruning (DP) adaptively zeros activations per input, but is coarse-grained (column-level) and keeps the dense weights resident in memory. The two are complementary—so a natural question is whether they can be jointly employed to push the speed–quality Pareto frontier further than either alone.

The perplexity evidence says yes. On Llama2-7B (WikiText), combined pruning attains its lowest perplexity at moderate total sparsity (~0.3–0.4) rather than at either extreme, and unified SP+DP (SPDP) sustains the lowest perplexity across all sparsity levels. The margin over static-only Wanda widens sharply beyond ~0.5—precisely the aggressive-sparsity regime where meaningful speedups become practically attainable.

PPL under combined pruning
(a) PPL under combined pruning
PPL across pruning methods
(b) PPL across pruning methods

Perplexity comparison across pruning strategies. (a) Combined pruning is best at moderate sparsity; (b) unified SP+DP (SPDP) yields the lowest perplexity across all sparsity levels, while static-only Wanda diverges at high sparsity.

Method

SPDP is built around a single design principle: make static and dynamic sparsity work through one representation and one execution model, so that both prefill and decode stay efficient. It maps three design challenges to three co-designed components.

  • Tiled-CBC format. A Tiled-Column-wise Bitmap Compressed layout stores unstructured statically-pruned weights as regular tiles with per-column bitmaps, so a single compressed representation is reused across prefill and decode without reformatting.
  • SPDP-spMspV kernel (decode). A CUDA-core sparse matrix × sparse vector kernel with Hybrid Activation-aware Dynamic Shared-Memory Bitmap Decoding (HAD-SMBD) that skips redundant activation columns at runtime, exploiting input-dependent sparsity in the bandwidth-bound decode phase.
  • SPDP-SpMM kernel (prefill). A Tensor-Core sparse matrix × matrix kernel that keeps prefill compute-efficient while reusing the same Tiled-CBC weights.
Overview of the Tiled-CBC format

(a) The Tiled-CBC format — nonzeros within each GTile are stored in column-major order using a hierarchical layout.

Design overview of SPDP

(b) Design overview — one compressed representation feeds the Tensor-Core SpMM kernel during prefill and the CUDA-core spMspV kernel during decode.

Kernel-Level Benchmark

We first benchmark the SPDP-spMspV decode kernel directly against six baselines—the Tensor-Core dense cuBLAS, cuSPARSE, Sputnik, SparTA, Flash-LLM, and the state-of-the-art SpInfer—with all speedups normalized to the Tensor-Core cuBLAS baseline, over LLM projection-matrix shapes.

On the A10G GPU, SPDP-spMspV achieves 1.88× average speedup over cuBLAS and 1.24× over SpInfer (up to 3.32× and 1.70×). On the L4 GPU, average speedups reach 2.11× and 1.37× (up to 3.52× and 2.52×). The advantage is largest in the moderate 30–50% static-sparsity regime targeted by practical LLM inference (shaded), where DP contributes input-dependent column skipping on top of static compression.

SPDP-spMspV kernel speedup vs. cuBLAS_TC across a 0-90% static-sparsity sweep on the A10G GPU

SPDP-spMspV kernel speedup over the Tensor-Core cuBLAS baseline on A10G, sweeping static sparsity 0–90% with DP fixed at 30%. The shaded band marks the moderate-sparsity regime targeted by LLM inference.

End-to-End Results

At matched perplexity, SPDP reaches up to 25% higher total sparsity than static-only pruning (SpInfer / Wanda) while delivering lower time-per-output-token (TPOT), across A10G, L4, and L40S GPUs. Because dynamic pruning is essentially “free” in quality, SPDP moves further along the efficiency–quality Pareto frontier than either static or dynamic pruning alone.

Sparsity ratio vs. TPOT across A10G, L4, and L40S for SPDP and SpInfer

Sparsity ratio vs. TPOT across GPUs. SPDP reduces TPOT while maintaining reasonable perplexity, whereas static-only pruning diverges beyond ~65% sparsity.

BibTeX

@article{kim2026spdp,
  author    = {Jinhyeok Kim and Yejoon Lee and Jaeyoung Do},
  title     = {Unified Static--Dynamic Pruning for Efficient LLM Inference},
  journal   = {Proceedings of the VLDB Endowment (PVLDB)},
  volume    = {19},
  number    = {11},
  pages     = {XXX--XXX},
  year      = {2026},
  doi       = {XX.XX/XXX.XX}
}