dMX: Differentiable Mixed-Precision Assignment for Low-Precision Floating-Point Formats

In three linesdMX is a differentiable mixed-precision quantization framework for learnable floating-point bit-width assignment across LLM layers. Tested on Llama, Qwen3, and SmolLM2 using the MXFP standard (Open Compute Project), it optimizes layer formats continuously then discretizes via annealing, outperforming KL-divergence heuristics on WikiText-2 and zero-shot reasoning benchmarks.

Summary generated by Claude — human-verified