Language Modeling on WikiText-103 (PPL, Bits, and Efficiency Delta)
446.8Perplexity (PPL)fp16
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| fp16Precision=fp162026.05 | 446.8 | 16 | 0 | |
| HQMQQuantization Setting=s96_r6 + Med3×2026.05 | 460.4 | 5.25 | 13.6 | |
| HQMQQuantization Setting=s192_r42026.05 | 469.5 | 4.04 | 22.7 | |
| HQMQQuantization Setting=s24_r6 + Med3×2026.05 | 542.2 | 4.78 | 95.4 | |
| HQMQQuantization Setting=s96_r42026.05 | 553.8 | 3.79 | 107 | |
| HQMQQuantization Setting=s48_r42026.05 | 688.1 | 3.54 | 241.3 | |
| HQMQQuantization Setting=s24_r32026.05 | 1,039 | 3.04 | 593 | |
| NaivePrecision=int42026.05 | 1,666 | 4 | 1,219 | |
| NaivePrecision=int32026.05 | 3,528 | 3 | 3,081 |