Inference Throughput for Abstractive Summarization on CNN/DailyMail
3.4Baseline Throughput (samples/s)DistilBART
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| DistilBARTArchitecture=12L-6L-1024, Architecture Type=encoder-decoder architecture2021.06 | 3.4 | 18.5 | 5.4 | |
| GPT2Architecture=0L-12L-768, Architecture Type=decoder-only architecture2021.06 | 3 | 16.7 | 5.5 | |
| ProphetNetArchitecture=12L-12L-1024, Architecture Type=encoder-decoder architecture2021.06 | 2.8 | 10.7 | 3.8 | |
| BARTArchitecture=12L-12L-1024, Architecture Type=encoder-decoder architecture2021.06 | 2.4 | 18.4 | 7.7 | |
| UniLMArchitecture=12L-0L-768, Architecture Type=encoder-only architecture2021.06 | 1.7 | 16.4 | 9.6 |