Entropy calibration chooses a quantization range by minimizing KL-divergence, preserving distribution shape rather than minimizing point-wise reconstruction error. That objective can fail on heavy-tailed activations: a tall peak near zero may lead the optimizer to shrink the range and clip sparse tails. section 4.3
Heavy-tailed activation failure
For the reported BERT activation distribution, entropy calibration chose a range of 0.7199 and clipped about 25% of values. Its MSE was 0.160087, versus 0.000600 for MSE-optimal calibration. section 4.3
CALIBRATION RESULTS ────────────────────────────────────────────────────────────── Method Range Scale Clipped MSE SNR (dB) ────────────────────────────────────────────────────────────── mse 10.5917 0.083399 0.0236% 0.000600 29.52 percentile 10.7027 0.084273 0.0101% 0.000604 29.49 absmax 11.5178 0.090691 0.0000% 0.000688 28.93 entropy 0.7199 0.005668 24.9989% 0.160087 5.26 ────────────────────────────────────────────────────────────── Best method: mse (MSE: 0.000600)
Compare the ranges, clipped fractions, and MSE values for entropy, MSE, percentile, and absmax calibration. section 4.3
Entropy calibration can work for roughly uniform or Gaussian distributions, but requires caution on heavy-tailed transformer activations. Check the observed distribution and validate clipping and reconstruction error before choosing it. section 4.3
