Symmetric quantization
Symmetric quantization maps values to integers using a zero-centered representation and a scale without an offset.
Quantization and Fast Inference
Symmetric quantization uses a zero-centered representation, typically forcing the weight zero-point to zero.
- How it is used
- Selects a zero-centered weight grid within hybrid INT8 inference to reduce correction overhead.
- How deep it goes
- Derives scale and zero-point behavior and implements calibration, quantization, dequantization, and corrected linear inference.
- Examples
- Compares symmetric and asymmetric ReLU grids and tests zero-point correction on shifted activations.
- What you can do after
- Builds a symmetric weight quantizer from scratch and evaluates its use in a hybrid quantized linear layer.
Read how Quantization and Fast Inference explains symmetric quantization →
CUDA for LLMs
Symmetric quantization maps values around zero to signed integer representations using a scale without an offset.
- How it is used
- Selects signed integer mappings for zero-centered weights in model quantization workflows.
- How deep it goes
- Works through scale selection and quantize/dequantize arithmetic, with elementwise CUDA kernels and deployment tradeoffs.
- Examples
- Quantizes four floating-point weights to INT8 and contrasts zero-centered weights with ReLU activations.
- What you can do after
- Build symmetric INT8 quantize and dequantize kernels from scratch and evaluate scheme choices against distribution and accuracy.
What to know first
- activationAn activation is a numerical value or tensor produced by a neural-network unit, layer, or operation during a forward computation.
- accuracyAccuracy is the degree to which a result agrees with a specified truth, reference, or criterion, commonly measured as the proportion of evaluated predictions that are correct.
- mean absolute errorMean absolute error is the arithmetic mean of the absolute differences between corresponding predicted and observed numerical values.
- mean squared errorMean squared error is the arithmetic mean of the squared differences between corresponding predicted and reference numerical values.
Related concepts
- asymmetric quantizationcontrasts with · Asymmetric quantization maps values to discrete levels using a scale and a zero point, allowing the represented range to be shifted relative to zero.
- hybrid consensusused by · Hybrid Consensus is a quantization strategy that applies symmetric quantization to weights and asymmetric quantization to activations.
- quantization scaleuses · A quantization scale is a multiplicative factor relating quantized values or steps to the numerical values they represent.
- weight-only quantizationused by · Weight-only quantization reduces the precision used to represent model weights while retaining higher precision for activations.
- affine quantizationcontrasts with · Affine quantization maps real values to discrete integer levels using a scale and an additive zero-point offset.
- quantizationkind of · Model quantization represents model parameters or computations with reduced-precision numerical values, often to lower storage or computational requirements.
Questions about symmetric quantization
What is symmetric quantization?
Symmetric quantization maps values to integers using a zero-centered representation and a scale without an offset.
Which Manning books explain symmetric quantization?
Quantization and Fast Inference and CUDA for LLMs.
What should I know before learning symmetric quantization?
Manning books that teach symmetric quantization usually introduce activation, accuracy, mean absolute error and mean squared error first.
What is the difference between symmetric quantization and asymmetric quantization?
Symmetric quantization maps values to integers using a zero-centered representation and a scale without an offset. Asymmetric quantization maps values to discrete levels using a scale and a zero point, allowing the represented range to be shifted relative to zero.
What is symmetric quantization a kind of?
Symmetric quantization is a kind of quantization. Model quantization represents model parameters or computations with reduced-precision numerical values, often to lower storage or computational requirements.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.

