Non-Uniform Quantization

The defining feature of non-uniform quantization is its variable step size between adjacent quantization levels. Unlike uniform quantization, where the step size (Δ) is constant, non-uniform schemes allocate smaller steps in regions of high value density (e.g., near zero for weights following a Gaussian distribution) and larger steps in the tails. This allows the finite set of quantized values to better approximate the original distribution, minimizing overall quantization error where it matters most.

The quantization levels are placed based on statistical analysis of the tensor's value distribution. Common methods include:

• Logarithmic Quantization: Uses a logarithmic scale, ideal for data with a heavy-tailed distribution (common in attention scores or certain activations).
• Companding (Compressing-Expanding): Applies a non-linear function (like μ-law or A-law) to the input, performs uniform quantization, and then applies the inverse function. This effectively allocates more levels to smaller amplitudes.
• K-Means Clustering: Uses clustering algorithms to place levels at the centroids of data clusters, directly minimizing the mean squared error for a given number of levels.

Non-uniform quantization excels when the data to be quantized has a non-uniform, often bell-shaped or long-tailed, distribution. Neural network weights and many activation outputs naturally follow such distributions (e.g., Gaussian, Laplacian). By concentrating quantization levels in the high-probability region, it captures more information per bit. This often results in higher accuracy compared to uniform quantization at the same bit-width, or allows for further bit-width reduction (e.g., to 4-bit) while maintaining acceptable accuracy.

The primary trade-off is computational complexity. Dequantization (mapping integer indices back to floating-point values) requires a non-linear lookup operation, as the step size is not constant. This lookup can be implemented via a look-up table (LUT). While LUT accesses are efficient, they introduce overhead not present in uniform quantization, where dequantization is a simple affine transformation (float_value = scale * int_value). This makes pure non-uniform quantization less favorable for core linear algebra operations (like GEMM) on standard hardware, though it can be efficient on custom Neural Processing Units (NPUs) with dedicated LUT support.

A widely studied non-uniform scheme is logarithmic quantization, where quantization levels follow a power-of-two progression. This is highly efficient because:

• Multiplication by a power-of-two becomes a simple bit-shift operation.
• The quantization scale itself becomes a power-of-two. It is particularly effective for quantizing weight tensors, where the dynamic range is large. Frameworks like TensorFlow Lite for Microcontrollers have explored log-based schemes for ultra-low-bit quantization on microcontrollers.

Non-uniform quantization is often used in conjunction with other advanced methods:

• Mixed-Precision Quantization: A model might use non-uniform (e.g., log) 4-bit quantization on weights and uniform 8-bit on activations.
• Per-Channel Quantization: Applying non-uniform schemes per-channel can yield even better fits for each filter's unique distribution.
• Quantization-Aware Training (QAT): QAT can learn optimal non-uniform level placements through gradient-based optimization using a Straight-Through Estimator (STE) to bypass the non-differentiable quantization function during backpropagation.

Uniform quantization is a scheme where the discrete quantization levels are evenly spaced across the represented range, resulting in a constant step size (Δ) between adjacent levels. This is the most common and hardware-friendly approach.

• Mechanism: The full range of values is divided into 2^b equal intervals, where b is the bit-width.
• Hardware Advantage: The constant step size allows for efficient integer arithmetic using simple scaling, as multiplication by the scale factor is a uniform operation.
• Trade-off: It can introduce significant error if the original data distribution is highly non-uniform, as many quantization levels may be wasted on regions with few data points.

Quantization error is the numerical distortion or loss of information introduced when mapping a continuous range of floating-point values to a finite set of discrete integer levels. It is the fundamental trade-off in any quantization scheme.

• Sources: Error arises from rounding (mapping to the nearest level) and clipping (saturating values outside the representable range).
• Measurement: Often quantified as the difference between the original and dequantized tensors, using metrics like Mean Squared Error (MSE).
• Non-Uniform Rationale: Non-uniform quantization aims to minimize this error for a given bit-width by placing more levels in high-probability regions of the value distribution.

The quantization scale and zero-point are the affine transformation parameters that map between floating-point and integer domains. They define the quantization scheme's granularity and offset.

• Scale (S): The ratio between the quantized integer step size and the floating-point range. A smaller scale allows finer granularity.
• Zero-Point (Z): An integer value that corresponds exactly to the floating-point zero. This enables efficient representation of asymmetric ranges and exact zero computation.
• Non-Uniform Context: In non-uniform schemes, the scale is not constant. The mapping function becomes non-linear, often implemented via a lookup table (LUT) rather than a simple affine transformation.

Calibration is the process of analyzing a representative dataset (the calibration set) to estimate the optimal parameters for quantization, such as value ranges, scales, and zero-points.

• Purpose: Determines how to best map the floating-point distribution to the discrete grid to minimize error.
• For Non-Uniform: Calibration is critical for identifying the regions of the value distribution with the highest density. Algorithms analyze histograms of tensor values to decide where to place more quantization levels.
• Methods: Common strategies include Min-Max (use observed min/max), Entropy Minimization (minimize KL divergence), and Percentile (use percentile estimates to exclude outliers).

Mixed-precision quantization is a technique that assigns different numerical precisions (bit-widths) to different layers, channels, or operators within a single model.

• Goal: To optimize the accuracy-efficiency trade-off by applying aggressive quantization to robust layers and preserving higher precision for sensitive ones.
• Relation to Non-Uniform: Both are forms of heterogeneous quantization. Mixed-precision varies bit-width across components, while non-uniform varies level density within a tensor. They can be combined.
• Automation: Often uses sensitivity analysis or reinforcement learning to automatically determine the optimal precision assignment per layer.

Quantization-Aware Training (QAT) is a model compression technique that simulates quantization effects during the training or fine-tuning process to improve the final accuracy of the low-precision model.

• Process: Inserts fake quantization nodes into the forward pass that round values but use full-precision gradients in the backward pass (via the Straight-Through Estimator).
• Advantage over PTQ: Allows the model weights to adapt to the quantization noise, typically yielding higher accuracy than Post-Training Quantization.
• Support for Non-Uniform: QAT frameworks can be extended to simulate non-uniform quantization during training, allowing the model to learn optimal representations for a non-linear quantizer.