Community post: KLD is misunderstood as a quantization quality metric

A Reddit post argues KLD is overrated as a metric for ranking quantized models. It says computing KLD on random text doesn't reflect inference-time distributions, and creators can easily inflate results by evaluating on calibration data. The author recommends generating data with the base model and ignoring prompt tokens.