Token Generation Efficiency on MMLU
22Attention Time (ms)TyphoonMLA
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| TyphoonMLAModel=DeepSeek-v3, Batch size=128, Prompt Type=Prompt C2025.09 | 22 | 50.1 | 2.56 | |
| TyphoonMLAModel=DeepSeek-v3, Batch size=128, Prompt Type=Prompt B2025.09 | 25.9 | 54 | 2.37 | |
| FlashMLAModel=DeepSeek-v3, Batch size=128, Prompt Type=Prompt C2025.09 | 26.9 | 55 | 2.33 | |
| FlashMLAModel=DeepSeek-v3, Batch size=128, Prompt Type=Prompt B2025.09 | 34.5 | 62.6 | 2.04 | |
| TyphoonMLAModel=DeepSeek-v3, Batch size=128, Prompt Type=Prompt A2025.09 | 58.1 | 86.3 | 1.48 | |
| FlashMLAModel=DeepSeek-v3, Batch size=128, Prompt Type=Prompt A2025.09 | 99.1 | 127.2 | 1.01 |