INT2 Quantization on OPT-125M
1.39Memory Cost (GB)OPTQ
Evaluation results
| Method | Links | |
|---|---|---|
| OPTQTarget=layer-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 1.39 | |
| Z-FOLDTarget=layer-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 1.39 | |
| OPTQTarget=layer-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 1.39 | |
| Z-FOLDTarget=layer-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 1.39 | |
| aespaTarget=attention-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 1.68 | |
| aespaTarget=attention-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 1.68 | |
| OmniQuantTarget=attention-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 1.94 | |
| OmniQuantTarget=attention-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 1.94 | |
| BRECQTarget=attention-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 3.39 | |
| BRECQTarget=attention-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 3.39 | |
| AffineQuantTarget=attention-wise reconstruction, Hardware=NVIDIA A100 GPU2024.02 | 3.47 | |
| AffineQuantTarget=attention-wise reconstruction, Hardware=single NVIDIA A100 GPU2024.02 | 3.47 |