Applied Compute adds QAD to improve quantized NVFP4 accuracy

Applied Compute researchers incorporated quantization-aware distillation (QAD) into their training platform to recover accuracy for quantized models, finding that large reductions in KL divergence against base models did not consistently correspond to recovered task accuracy. The authors observed that preserving high precision for attention components while quantizing MLPs significantly reduced KL, and they experimented with normal versus deployment-informed QAD objectives to mitigate training-inference mismatches. Open article →