End-to-end decode throughput on 128K Context Length
108Throughput (tok/s)Llama-3.1-8B
Evaluation results
| Method | Links | |
|---|---|---|
| Llama-3.1-8BKV Representation=SphKV2026.05 | 108 | |
| GPT-ossKV Representation=SphKV2026.05 | 102.4 | |
| Qwen2.5-14BKV Representation=SphKV2026.05 | 94.8 | |
| Llama-3.1-8BKV Representation=Dense2026.05 | 62.9 | |
| GPT-ossKV Representation=Dense2026.05 | 59.9 | |
| Qwen2.5-14BKV Representation=Dense2026.05 | 55.7 |