Throughput Evaluation on NLP variable-length
5.03Throughput (M tok/s)ELSA-FP16
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ELSA-FP16Precision=FP162026.04 | 5.03 | 424 | |
| ELSA-TurboPrecision=TF322026.04 | 2.643 | 175.3 | |
| ELSA-StrictPrecision=FP322026.04 | 2.004 | 108.7 | |
| SDPA-FP32-memPrecision=FP322026.04 | 0.96 | — |