llm-compression
Everything on Ground Truth tagged “llm-compression” — 1 item.
Post-Training Quantization: Calibration, GPTQ and Why Rotations Help Lesson
Post-training quantization shrinks a trained model to 4 bits using a small calibration set; methods like GPTQ correct rounding errors layer by layer and Hadamard rotations spread outlier values so they quantize well.