GFLOPs per token on E2E for Natural Language Generation
4.65GFLOPs per tokenLTE
Evaluation Results
| Method | Links | |
|---|---|---|
| LTEBackbone=LLaMA-7B, Permitted Quality Drop=ROUGE-L <= 0.052024.02 | 4.65 | |
| Deja VuBackbone=LLaMA-7B, Permitted Quality Drop=ROUGE-L <= 0.052024.02 | 6.42 | |
| R-LLaMA+MoEBackbone=LLaMA-7B, Permitted Quality Drop=ROUGE-L <= 0.052024.02 | 7.39 | |
| FullBackbone=LLaMA-7B, Permitted Quality Drop=ROUGE-L <= 0.052024.02 | 12.06 |