PracticeConfig Configuration Reference

1. Configuration Overview

Complete best-practice quantization task configuration: metadata + task (task ≡ apiversion + spec).

Item	Content
Config Class	PracticeConfig
Source	interface.py

2. Parameter List

2.1 PracticeConfig

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
metadata	object	Optional	See the nested configuration default	—	Quantization configuration metadata (config_id/score/label/verified_*)	This page §2.2
task	object	Optional	See the nested configuration default	—	Quantization task description (apiversion + spec)	This page §2.3

Configuration Constraints

• Preprocessing: the old format {apiversion, metadata, spec, ...} becomes {metadata, task:{apiversion, spec, ...}}.
• Strips the aggregated task. prefix so that error paths match the spec. paths in the YAML.

2.2 Metadata

Metadata describing a quantization configuration; the label and scenario fields index it for the best-practice library.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
config_id	string	Optional	Unknown	—	Quantization configuration ID, for example 'Qwen3-32B W8A8'	None
score	float	Optional	100.0	—	Quantization configuration score, used for sorting; must be >= 0	None
label	object	Optional	{}	—	Quantization configuration labels, used for filtering, for example: w_bit (int, weight quantization bit width), a_bit (int, activation quantization bit width), is_sparse (bool, whether sparse quantization is used), kv_cache (bool, whether KV Cache quantization is enabled), fa_quant (bool, whether FA quantization is enabled), fa_bit (int, FA quantization bit width)	None
verified_model_types	list[string]	Optional	[]	—	List of verified model types, for example ['LLaMa3.1-70B', 'Qwen2.5-72B']	None
verified_tags	object	Optional	{}	—	Verified scenario tags: keys are model types, values are lists of scenario tags (each scenario is a set of tags, such as ['MindIE','Atlas_A2_Inference'])	None

Configuration Constraints

• None.

2.3 BaseQuantConfig

Derived Classes

Config Class	Description	Document
ModelslimV1QuantConfig	modelslim_v1 quantization task configuration, located at the YAML root.	modelslim_v1 Configuration Reference
ModelslimConvertQuantConfig	modelslim_convert quantization (weight conversion) task configuration, located at the YAML root.	modelslim_convert Configuration Reference

3. Complete Configuration Reference

metadata:
  config_id: Qwen3-32B W8A8
  label:
    w_bit: 8
    a_bit: 8
    is_sparse: false
    kv_cache: false
  verified_model_types:
  - Qwen3-32B
task:
  apiversion: modelslim_v1
  spec:
    process:
    - type: linear_quant
      qconfig:
        act:
          dtype: int8
          scope: per_tensor
          symmetric: false
          method: minmax
        weight:
          dtype: int8
          scope: per_channel
          symmetric: true
          method: minmax
      include:
      - '*'
      exclude: []
    save:
    - type: ascendv1_saver
      part_file_size: 4
    dataset: mix_calib.jsonl