Research map / Architectures and efficiency / Pruning and efficient tuning
MACKO: Sparse matrix-vector multiplication for low sparsity
Vladimír Macko, Vladimír Boža
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MACKO sparse format instead of dense cuBLAS at 30% sparsity | Llama2-7b generation · Llama2-7b Wanda, RTX 4090 | tokens/s: 66.49 -> 73.88 | Improved |
| use MACKO instead of dense at 50% sparsity | Llama2-7b generation · Llama2-7b Wanda, RTX 4090 | tokens/s: 66.53 -> 98.60 | Improved |
| use MACKO instead of dense at 90% sparsity | Llama2-7b generation · Llama2-7b Wanda, RTX 4090 | tokens/s: 66.48 -> 255.01 | Improved |
| use MACKO instead of dense at 20% sparsity | Llama2-7b generation · Llama2-7b Wanda, RTX 4090 | tokens/s and size: 66.55 -> 66.36 toks/sec; 13.59 -> 13.77 GB | Hurt |
| use MACKO model size at 90% sparsity vs dense | Llama2-7b · Llama2-7b Wanda, RTX 4090 | model size (GB): 13.59 -> 2.67 | Improved |
| raise sparsity from 50% to 70% with MACKO SpMV | MACKO-SpMV 12288x12288 · RTX 4090 | speedup over cuBLAS: 1.3x -> 1.96x | Improved |
| raise sparsity to 95% with MACKO SpMV | MACKO-SpMV 12288x12288 · RTX 4090 | speedup over cuBLAS: 5.62x at 95% sparsity, but outperformed by Sputnik | Mixed |
| not pruning embedding layers (Wanda) limits end-to-end gain at 90% sparsity | Llama2-7b with MACKO · Llama2-7b 90% sparsity | speedup and memory reduction: 2% of original size -> 10% of 90%-sparse model | Hurt |
| use b_delta=1 index width (ablation) | MACKO format · storage | memory vs dense: improves over dense from 7% sparsity | Mixed |
Closest papers by meaning
- R-Sparse: Rank-Aware Activation Sparsity for Efficient LLM Inference (ICLR 2025 · similarity 0.89)
- MUSTAFAR: Promoting Unstructured Sparsity for KV Cache Pruning in LLM Inference (NeurIPS 2025 · similarity 0.89)
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.89)
- DuoGPT: Training-free Dual Sparsity through Activation-aware Pruning in LLMs (NeurIPS 2025 · similarity 0.88)
- Týr-the-Pruner: Structural Pruning LLMs via Global Sparsity Distribution Optimization (NeurIPS 2025 · similarity 0.88)
- Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
