RAG Prefill Acceleration on Llama 8B 64K Context Length
2TTFT (s)SIFT
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SIFTMode=SIFT2026.06 | 2 | 13.76 | 0.073 | 27.47 | 54.86 | |
| RecomputeMode=Recompute2026.06 | 2.92 | 17.94 | 0.056 | 52.31 | 152.53 | |
| CacheBlendMode=CacheBlend2026.06 | 4.3 | 10.75 | 0.093 | 46.2 | 198.5 |