awq Configuration Reference

1. Configuration Overview

AWQ (Activation-aware Weight Quantization) processor configuration.

Item	Content
Config Class	AWQProcessorConfig
Source	processor.py

2. Parameter List

2.1 AWQProcessorConfig

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
type	string	Optional	awq	awq	Processor type, fixed to awq.	None
weight_qconfig	object	Required	None	—	Weight quantization configuration (QConfig); required. See QConfig Configuration Reference.	This page §2.2
enable_subgraph_type	list[any]	Optional	['norm-linear', 'linear-linear', 'ov', 'up-down']	—	List of subgraph types for AWQ; the default is norm-linear, linear-linear, ov, up-down.	None
n_grid	int	Optional	20	—	Number of grids for the grid search, used to find the optimal quantization scale/clipping point; must be greater than 0.	None
include	list[string] / null	Optional	null	—	Module name patterns to include; if unset, all modules match.	None
exclude	list[string] / null	Optional	null	—	Module name patterns to exclude; takes precedence over include.	None

Configuration Constraints

• None.

2.2 QConfig

Describes how a single tensor (weight or activation) is quantized.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
dtype	string	Required	None	float, int8, int4, mxfp8, mxfp4, fp8_e4m3	Quantization data type, such as int8, int4, mxfp8, mxfp4, fp8_e4m3; float means the tensor is not quantized.	None
scope	string	Required	None	per_tensor, per_channel, per_group, per_block, per_token, pd_mix, per_head, dual_scale, per_channel_neg_offset	Quantization granularity, that is, the range over which scale/zero_point is computed. For different quantization targets, the available scope values usually differ: Activation: per_tensor (a single scale for the whole tensor), per_token (per attention head), per_block (per fixed block), pd_mix (mixed Prefill/Decode granularity); Weight: per_channel (per channel), per_group (per group), per_block (per fixed block), dual_scale (dual scale); FA3 attention: per_head (per attention head), per_token (per attention head).	None
symmetric	bool	Required	None	—	Whether to use symmetric quantization. Symmetric quantization stores only the scale; asymmetric quantization additionally stores a zero_point. Availability depends on the dtype/scope combination.	None
method	string	Required	None	—	Quantization parameter estimation algorithm, such as minmax, mse_round, histogram, ssz, none, gptq, and more. Algorithms belong to one of two quantizers by tensor type: the activation quantizer supports minmax, histogram, and none (together with float); the weight quantizer supports minmax, mse_round, ssz, gptq, and more. The available values also depend on the dtype/scope/symmetric combination; none means the parameters are not estimated.	None
ext	object	Optional	{}	—	Quantizer extension parameters, determined by method and the quantizer implementation (for example percdamp/group_size for gptq); an empty object means no extension parameters.	None

Configuration Constraints

• None.

3. Complete Configuration Reference

apiversion: modelslim_v1
spec:
  process:
  - type: awq
    weight_qconfig:
      dtype: int8
      scope: per_channel
      symmetric: true
      method: minmax
      ext: {}
    enable_subgraph_type:
    - norm-linear
    - linear-linear
    - ov
    - up-down
    n_grid: 20