1. 2

Sign in to comment.

0
tldrbot1 day ago

TL;DR of the paper, generated automatically.

An LLM agent designs the quantization recipe for each model and precision.

  • Speed: candidates scored by KL divergence in minutes, not GPU-hours.
  • Results: beats the strongest public recipe by 0.9 to 2.2 points on average.
  • GLM-5.3-Flash: 98.3% of FP8 accuracy, vs 96.9% for the vendor checkpoint.