Layer Pruning Efficiency Evaluation on LLaVA-NeXT 8B
18.1Peak VRAM (GB)TopoVLM
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| TopoVLMSparsity=50%2026.04 | 18.1 | 5.7 | 1 | 60.3 | 1.75 | |
| TAMPSparsity=50%2026.04 | 18.3 | 16.8 | 5 | 63.4 | 1.66 | |
| WandaSparsity=50%2026.04 | 18.5 | 18.9 | 6 | 64.9 | 1.62 | |
| SparseGPTSparsity=50%2026.04 | 18.7 | 22.3 | 8 | 65.8 | 1.6 | |
| LLM-PrunerSparsity=50%2026.04 | 19.2 | 14.6 | 5 | 67.1 | 1.57 | |
| Full ModelSparsity=0%2026.04 | 31.8 | — | — | 105.4 | 1 | |
| LaCoSparsity=50%2026.04 | — | 17.4 | 5 | — | — |