: Neural network weights typically follow a normal distribution. NF4 concentrates its 16 "bins" where most weights exist (near zero), minimizing rounding errors.
: A process that quantizes the quantization constants themselves to save additional memory.
In the context of computer science and machine learning, refers to 4-bit NormalFloat , a specialized quantization data type introduced in the seminal paper QLoRA: Efficient Finetuning of Quantized LLMs by Tim Dettmers et al. (2023). 📄 Core Concept: The QLoRA Paper NF4.rar
: To reduce the memory footprint of LLMs (like Llama) enough to fit on a single GPU (e.g., a 24GB RTX 3090) while maintaining full 16-bit performance.
: An information-theoretically optimal data type for normally distributed weights. It uses 16 quantization levels based on the quantiles of a standard normal distribution. : Neural network weights typically follow a normal
: Recent research (April 2026) has further optimized this by creating Fast NF4 Dequantization Kernels that achieve 2.0–2.2× speedups on NVIDIA GPUs. ⚠️ Alternative Interpretation
: RNF4 mediates the degradation of the PML-RARα fusion protein. In the context of computer science and machine
The paper explains why NF4 is superior to standard 4-bit integers (Int4) or floating-point (Float4) formats: