Graph Quantizer

Graph Quantizer (GQ) reduces model precision at the intermediate representation (NPIR) level using calibration data. It converts floating-point weights and activations to integers with lower bit widths. This process enables faster inference and smaller model size while keeping accuracy degradation within acceptable bounds.

Overview

Graph Quantize runs after Graph Optimize and operates on the optimized NPIR. It uses calibration data to measure activation ranges for each layer. It then assigns quantization parameters (scale, zero-point) to each operator.

Pipeline position

GQ supports standard quantization and Automatic Mixed Precision (AMP). AMP automatically assigns different bit widths to different layers.

Configuration

Parameters

Parameter	Required	Description	Accepted values
scheme	Yes	Weight and activation bit width	W8-A8, W8-A8_dynamic, W4-A8_dynamic

Hardware availability

Supported schemes vary by target hardware backend. The device that you configure at project creation determines the available options.

Backend	Supported schemes
XNNPACK	W8-A8, W8-A8_dynamic, W4-A8_dynamic
Ethos-U	W8-A8, W8-A16

Choosing a scheme

• Start with W8-A8; it is the most stable and widely tested combination.
• Use W8-A8_dynamic when you need dynamic activation quantization.
• Use W4-A8_dynamic for aggressive INT4 weight compression. This scheme is recommended for large language models (LLMs).

YAML configuration

steps:
- graph_optimize: null
- graph_quantize:
  - scheme: W8-A8     # Available (XNNPACK): W8-A8, W8-A8_dynamic, W4-A8_dynamic / (Ethos-U): W8-A8, W8-A16
    automatic_mixed_precision: null  # Available: GREEDY

Advanced overrides (run.detail.yaml)

For fine-grained control over calibration behavior, create a run.detail.yaml file in the same experiment directory.

run.detail.yaml

detail:
  graph_quantize[0]:
    weight:
      granularity: CHANNEL  # Available: CHANNEL, TENSOR
      calibration_method: MINMAX  # Available: MINMAX, PERCENTILE
    activation:
      granularity: TENSOR  # Available: TENSOR
      calibration_method: PERCENTILE  # Available: MINMAX, PERCENTILE
      percentile: 99.9
    parameters:
      calibration:
        num_samples: -1

note

parameters.calibration is not available when using W8-A8_dynamic or W4-A8_dynamic schemes. Dynamic activation quantization determines activation ranges at runtime and does not use a separate calibration pass.

Calibration methods

Method	Behavior	When to use
MINMAX	Maps the absolute max float value to the max integer	Default: safe for most models
PERCENTILE	Clips top and bottom outliers before mapping	Use when activations have occasional spikes

See run.detail.yaml for the full Mixed Precision configuration reference.

Calibration data

GQ requires calibration data to measure activation ranges for each layer. Use a preset dataset selected during np projects create.

Log output

============================================================
[2/4] Graph Quantize
============================================================
2026-08-21 16:08.22 [INFO ] [Workflow] Graph Quantize started

2026-08-21 16:10.15 [DONE ] [Workflow] Graph Quantize completed (1m 52s)

Performance benefits

• Reduced model size: INT8 weights are approximately 4 times smaller than FP32, and INT4 weights are approximately 8 times smaller
• Faster inference: Integer arithmetic executes significantly faster than float on most hardware
• Controlled accuracy trade-off: Calibration and AMP enable fine-grained control over the balance between accuracy and performance