Weight Reconstruction Fidelity on DeepSeek-V3 Weights
0Weight ΔW L2 DistancePost-trained
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Post-trainedPrecision=BF162026.03 | 0 | 100 | 1 | |
| AbsMaxPrecision=FP8, Quantization Granularity=block2026.03 | 48,641 | 54.54 | 0.239 | |
| AbsMaxPrecision=FP8, Quantization Granularity=channel2026.03 | 51,790 | 62.48 | 0.261 |