dMX: Differentiable Mixed-Precision Assignment for Low-Precision Floating-Point Formats
Researchers have developed dMX, a novel differentiable framework for optimizing the bit-width of floating-point formats in large language models. This method allows for learnable, per-layer bit-width assignments, moving beyond uniform quantization to improve both accuracy and performance. Experiments on models like Llama and Qwen3 demonstrate that dMX can achieve better trade-offs between model quality and deployment efficiency compared to existing heuristics. AI
IMPACT Enables more efficient deployment of large language models by optimizing their precision.