Delta Accuracy on RULER Long-context Understanding
6.9Delta AccuracyCertified Quantized Attention
Evaluation Results
| Method | Links | |
|---|---|---|
| Certified Quantized AttentionContext Length=8K, n=20 slices, Backbone=LLaMA 3.1-8B, Key Precision=INT8 (asymmetric), Value Precision=INT4 (g=16)2026.05 | 6.9 | |
| Certified Quantized AttentionContext Length=32K, n=20 slices, Backbone=LLaMA 3.1-8B, Key Precision=INT8 (asymmetric), Value Precision=INT4 (g=16)2026.05 | 1.9 | |
| Certified Quantized AttentionContext Length=64K, n=20 slices, Backbone=LLaMA 3.1-8B, Key Precision=INT8 (asymmetric), Value Precision=INT4 (g=16)2026.05 | 1.5 | |
| Certified Quantized AttentionContext Length=128K, n=20 slices, Backbone=LLaMA 3.1-8B, Key Precision=INT8 (asymmetric), Value Precision=INT4 (g=16)2026.05 | 0.4 |