INT2 Quantization on LLaMA-30B
18.59Memory Footprint (GB)OPTQ
Evaluation results
| Method | Links | |
|---|---|---|
| OPTQTarget=layer-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 18.59 | |
| Z-FOLDTarget=layer-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 18.59 | |
| OPTQTarget=layer-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 18.59 | |
| Z-FOLDTarget=layer-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 18.59 | |
| OmniQuantTarget=attention-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 24.53 | |
| OmniQuantTarget=attention-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 24.53 | |
| AffineQuantTarget=attention-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 38.59 | |
| AffineQuantTarget=attention-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 38.59 | |
| aespaTarget=attention-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 43 | |
| aespaTarget=attention-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 43 |