notesum.ai
Published at November 5Change Is the Only Constant: Dynamic LLM Slicing based on Layer Redundancy
cs.CL
cs.AI
cs.LG
I.2.7; I.2.0
Released Date: November 5, 2024
Authors: Razvan-Gabriel Dumitru1, Paul-Ioan Clotan2, Vikas Yadav3, Darius Peteleaza4, Mihai Surdeanu1
Aff.: 1University of Arizona; 2Università di Bologna; 3ServiceNow AI; 4Lucian Blaga University of Sibiu

| Model | Technique | Pruned | Piqa | Hellaswag | Winogrande | Arc Easy | Wikitextv2 | Average |
|---|---|---|---|---|---|---|---|---|
| Acc. () | Acc. () | Acc. () | Acc. () | Perplexity () | Acc. () | |||
| Llama 3-8B | SliceGPT | 30% | 59.3% | 37.2% | 56.4% | 42.9% | 13.37 | 49.0% |
| 35% | 57.7% | 34.1% | 54.3% | 39.3% | 16.58 | 46.4% | ||
| 40% | 57.0% | 32.4% | 51.8% | 35.9% | 20.69 | 44.3% | ||
| Dynamic Slicing | 30% | 60.4% | 38.4% | 58.0% | 42.4% | 12.96 | 49.8% | |
| 35% | 58.4% | 36.3% | 57.2% | 39.3% | 15.64 | 47.8% | ||
| 40% | 58.1% | 34.0% | 54.4% | 36.8% | 19.11 | 45.8% | ||
| Mistral-7B | SliceGPT | 30% | 62.6% | 38.0% | 59.7% | 51.1% | 8.87 | 52.9% |
| 35% | 58.5% | 35.9% | 57.6% | 42.8% | 10.80 | 48.7% | ||
| 40% | 57.1% | 33.6% | 54.1% | 38.2% | 13.33 | 45.8% | ||
| Dynamic Slicing | 30% | 63.1% | 38.6% | 60.2% | 51.7% | 8.76 | 53.4% | |
| 35% | 58.5% | 34.9% | 55.7% | 45.8% | 10.38 | 48.8% | ||
| 40% | 57.9% | 31.9% | 54.1% | 40.1% | 12.62 | 46.0% |