Record#
本周工作#
- [x]
模型 Qwen2.5-omni-3B,bf16
- 2 fps
| Model | fps | overall_accuracy |
|---|---|---|
| Full tokens | 2 | 46.4 |
| V(0.5)+A(0.3), new | 2 | 41.9 |
| V(0.5)+A(0.2), new | 2 | - |
| V(0.2)+A(0.2), new | 2 | 44.7 |
| V(0.5)+A(0.3), new+llm | 2 | 25.1 |
暂时不继续做 2 fps 的实验,还是有部分样本爆显存,还没找原因
- 0.5 fps
| Model | RetainedRatio | fps | overall_accuracy |
|---|---|---|---|
| Full tokens | 100% | 0.5 | 45.5 |
| V(0.6)+A(0.3), omnizip改 | - | 0.5 | 44.9 |
| V(0.5)+A(0.5), omnizip改 | - | 0.5 | 44.7 |
| V(0.5)+A(0.3), omnizip改 | - | 0.5 | 45.1 |
目前欠缺的实验,
- 目前 baseline 的效果和其 RetainedRatio
- 目前方法的和效果 RetainedRatio
- 所有实验的 RetainedRatio