Quantizing deep neural networks is essential for squeezing massive models onto resource-constrained hardware. Traditionally, this process maps high-precision floating-point weights to lower-precision integers, such as 8-bit, 4-bit, or 2-bit representations.
However, this conventional approach suffers from a fundamental limitation: integer rigidity. Hardware memory capacities are continuous (e.g., a 24GB GPU), but integer quantization forces model sizes to scale down in massive, discrete steps. As noted by Zihan Chen et al. in Channel-Wise Mixed-Precision Quantization for Large Language Models, this lack of granularity forces developers into an uncomfortable compromise: either use a larger integer bit-width that exceeds the device's memory budget, or drop down an entire integer bit, sacrificing substantial model accuracy while leaving valuable hardware memory completely unutilized.
Liulu He et al. in LiftQuant: Continuous Bit-Width LLM via Dimensional Lifting and Projection describe this mismatch as a "deployment gap." For instance, fitting a 70-billion parameter model perfectly into 24GB of VRAM might require an average precision of exactly 2.4 bits per weight. Traditional quantization frameworks simply cannot do this.
Fractional bit-width quantization breaks this integer barrier, allowing models to achieve non-integer average bit-widths (such as 2.2 or 2.4 bits) to perfectly "right-size" a model for any given hardware target.
What is a Fractional Bit?
To build intuition, we must first clarify what a "fractional bit" means in practice. A physical hardware register cannot store 2.4 bits of data. Instead, fractional quantization works by decoupling the average stored memory footprint from the underlying computational primitives.
By allowing the average bit-rate across the network to be a non-integer, the model can smoothly traverse the accuracy-compression Pareto frontier. This granularity allows model compression to approach theoretical distortion limits much more closely than rigid integer schemes, minimizing the performance loss typically caused by rounding bit assignments down to the nearest integer (Deokjae Lee et al. in Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment).
Researchers achieve these fractional averages through two primary strategies: granular mixed-precision allocation and mathematical reformulations.
Mechanism 1: Granular Mixed-Precision Allocation
The most direct way to achieve a fractional average bit-width is to mix different integer bit-widths at a highly granular level. If half of a model's channels are quantized to 2 bits and the other half to 3 bits, the network effectively operates at a fractional average of 2.5 bits.
Early foundations for this flexibility were established by libraries like BMXNet (Haojin Yang et al. in BMXNet: An Open-Source Binary Neural Network Implementation Based on MXNet) and entropy-driven frameworks (Eunhyeok Park et al. in Weighted-Entropy-Based Quantization for Deep Neural Networks), which proved that multi-bit representations could be tailored to specific hardware constraints.
Modern fractional quantization refines this by determining exactly where to allocate higher precision based on sensitivity:
- Channel and Layer-Wise Mixing: Zihan Chen et al. allocate different bit-widths to individual weight channels based on activation sensitivity. Similarly, Razvan-Gabriel Dumitru et al. in Variable Layerwise Quantization: A Simple and Effective Approach to Quantize LLMs identify critical layers to keep at higher precision while aggressively compressing less important ones, hitting precise averages like 2.85 bits.
- Filter and Neuron-Wise Mixing: Wenhao Sun et al. in Class-based Quantization for Neural Networks assign bit-widths at the filter level based on their importance to specific dataset classes.
- Dynamic Block Offloading: Moving beyond static allocation, Xinghao Wang et al. in BitStack: Any-Size Compression of Large Language Models in Variable Memory Environments decompose models into residual blocks, allowing pieces of the model to be dynamically loaded or offloaded on the fly to meet megabyte-level memory budgets.
Mechanism 2: Mathematical and Structural Reformulations
An alternative to mixing discrete integer layers is to reformulate how weights are represented mathematically, embedding fractional scaling directly into the quantization operators.
Mantissa Sharing
For floating-point formats, the authors of AMS-Quant: Adaptive Mantissa Sharing for Floating-Point Quantization (Mengtao Lv et al.) demonstrate that non-integer bit-widths can be achieved by allowing multiple weights to share their least significant mantissa bits. This provides a continuous knob for compression without requiring entirely new hardware formats for every fractional step.
Dimensional Lifting and Projection
One major engineering hurdle for fractional quantization is that standard hardware accelerators require specialized compute kernels for every supported bit-width. To bypass this, Liulu He et al. developed a "lift-then-project" mechanism in LiftQuant and Lifted Uniform Quantization for Extreme Low-bit Large Language models. By projecting weights into a higher-dimensional space before uniform quantization, they achieve the mathematical precision of fractional bit-widths while remaining fully compatible with standard, fast uniform quantization hardware operators.
Redundant Spaces and Factorization
Other techniques map weights into alternative mathematical spaces that are inherently robust to noise. Harshavardhan Adepu et al. in FrameQuant: Flexible Low-Bit Quantization for Transformers leverage a "Fusion Frame" space from signal processing to enable stable, fractional bit-rates like 2.2 bits.
Similarly, Vladimír Boža et al. in Addition is almost all you need: Compressing neural networks with double binary factorization decompose weight matrices into products of binary matrices paired with scaling vectors, replacing energy-intensive multiplications with simple additions. For sub-1-bit ranges, Dongsoo Lee et al. introduced FleXOR: Trainable Fractional Quantization, which uses differentiable XOR gates to learn encrypted weight representations at fractional widths as low as 0.6 bits per weight.
Additionally, specialized hardware design environments, such as those proposed by R. Kanda et al. in Design Environment of Quantization-Aware Edge AI Hardware for Few-Shot Learning, provide fine-grained control over both the integer and fractional parts of fixed-point representations, ensuring software simulations perfectly match hardware deployment.
Why It Works: The Intuition of Outlier Preservation
Why does moving from a strict 2.0-bit model to a 2.2-bit model yield disproportionately large improvements in model perplexity?
The core intuition lies in the non-uniform sensitivity of neural networks. In large language models and deep networks, a tiny fraction of weights or channels (often called "emergent features" or "outliers") carry a massive portion of the model's representational capacity.
Uniform 2-bit Quantization (Rigid):
[ Normal Weight (2-bit) ] [ Critical Outlier (2-bit - HEAVILY DISTORTED) ] -> High Perplexity Loss
Fractional 2.2-bit Quantization (Flexible):
[ Normal Weight (2-bit) ] [ Critical Outlier (3-bit - PRESERVED) ] -> Low Perplexity Loss
If you force a strict 2-bit constraint uniformly across the model, these critical outlier features are heavily distorted, causing catastrophic accuracy drops. Conversely, if you upgrade the entire model to 3 bits, the model size inflates past your hardware capacity.
Fractional quantization works because it allows the framework to be selective. By utilizing fractional bit-widths, the system can allocate a higher bit-depth (like 3-bit or 4-bit) exclusively to the fragile, high-importance parameters, while keeping the rest of the network at an aggressive 2-bit baseline. As argued by Runsheng Bai et al. in SKIM: Any-bit Quantization Pushing The Limits of Post-Training Quantization, this granular flexibility ensures that we never sacrifice performance by rounding down to the nearest integer bit-width, offering a perfect mathematical bridge between extreme compression and model quality.
Key Foundational Work
Weighted-Entropy-Based Quantization for Deep Neural Networks (2017) — contributed 2 concept elements to this direction.
FleXOR: Trainable Fractional Quantization (2020) — contributed 1 concept elements to this direction.
BMXNet: An Open-Source Binary Neural Network Implementation Based on MXNet (2017) — contributed 1 concept elements to this direction.