INT2 Quantization on LLaMA-13B
12.34Memory Cost (GB)OPTQ
Evaluation Results
| Method | Links | |
|---|---|---|
| OPTQTarget=layer-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 12.34 | |
| Z-FOLDTarget=layer-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 12.34 | |
| OPTQTarget=layer-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 12.34 | |
| Z-FOLDTarget=layer-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 12.34 | |
| OmniQuantTarget=attention-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 17.02 | |
| OmniQuantTarget=attention-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 17.02 | |
| AffineQuantTarget=attention-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 27.1 | |
| AffineQuantTarget=attention-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 27.1 | |
| aespaTarget=attention-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 29.27 | |
| aespaTarget=attention-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 29.27 |