Configuration Reference Navigation

This directory is the field-level reference for msModelSlim configuration (YAML). It is organized as: choose a protocol first, then look up the type. For the usage workflow and end-to-end steps, see One-Click Quantization Usage Guide; for the task protocol overview, see Quantization Task Configuration Reference Navigation; for the auto-tuning protocol overview, see Auto Tuning Configuration Reference Navigation.

1. How to Use This Documentation Set

1. Choose the protocol: the apiversion at the YAML root determines the structure of spec. Select the protocol page by task scenario (for an overview, see Quantization Task Configuration Reference Navigation). Field values and the available process / save subtypes differ between protocols; the corresponding task page is authoritative.
2. Look up processors: each item of spec.process[] is dispatched by type. Use the type to find the corresponding page in the processors table below; spec.prior[].process references the same processors.
3. Look up save formats: each item of spec.save[] is dispatched by type. Use the save formats table below to find the corresponding page; the VLM / multimodal generation protocols accept only the AscendV1 and MindIE types.
4. Look up tuning: for the top-level auto-tuning structure, see Auto Tuning Configuration Reference Navigation; for the fields of strategy and evaluation, see the auto tuning table below.

This page, Quantization Task Configuration Reference Navigation, and Auto Tuning Configuration Reference Navigation are hand-written navigation pages; the remaining configuration pages are generated from source Pydantic annotations by skills/docs-management/scripts/gen_config_api_docs.py. To change a field description, edit the source annotations such as Field(description=) and regenerate; do not edit pages that carry a generated-by marker directly.

2. Task Configuration (Selected by apiversion)

To look up fields by protocol, see Quantization Task Configuration Reference Navigation. The table below is a quick entry point:

Configuration	apiversion	Description
modelslim_v1	modelslim_v1	Default protocol for large language model quantization, including runner/process/save/dataset.
multimodal_vlm_modelslim_v1	multimodal_vlm_modelslim_v1	Quantization protocol for multimodal understanding (VLM) models.
multimodal_sd_modelslim_v1	multimodal_sd_modelslim_v1	Quantization protocol for multimodal generation (DiT/SD) models, including dump and inference parameters.
modelslim_convert	modelslim_convert	Pure weight conversion protocol, used for weight renaming, structure replacement, and format conversion.
practice_config	modelslim_v1 (inside task)	Best-practice configuration: metadata + task, referenced by auto-tuning strategies.

3. Processors (spec.process[] Dispatched by type)

type	Configuration	Purpose
adapt_rotation	adapt_rotation	Adaptive rotation; suppresses outliers.
autoround_quant	autoround_quant	AutoRound quantization training.
awq	awq	AWQ activation-aware weight quantization.
binary_analysis	binary_analysis	Binary (quantized/non-quantized) sensitivity analysis.
binary_operator_layer_wise	binary_operator_layer_wise	Layer-wise sensitivity analysis.
binary_operator_model_wise	binary_operator_model_wise	Model-level sensitivity analysis.
dynamic_cache	dynamic_cache	KV cache dynamic quantization.
fa3_quant	fa3_quant	FA3 attention quantization.
flatquant	flatquant	FlatQuant quantization training.
flex_awq_ssz	flex_awq_ssz	FlexAWQSSZ smooth quantization.
flex_smooth_quant	flex_smooth_quant	FlexSmoothQuant smooth quantization.
float_sparse	float_sparse	Floating-point sparsity.
group	group	Combiner processor; shares the inference process.
iter_smooth	iter_smooth	Iterative smoothing.
kv_smooth	kv_smooth	KV cache smoothing.
linear_quant	linear_quant	Linear layer quantization, including qconfig (activation/weight).
load	load	Module load/unload.
oasq	oasq	OASQ outlier-aware smooth quantization.
online_quarot	online_quarot	Online QuaRot rotation.
quarot	quarot	Offline QuaRot rotation.
saver	saver	Unified save processor.
smooth_quant	smooth_quant	SmoothQuant smooth quantization.
svd_res	svd_res	SVD low-rank residual compensation.
trainable_linear_quant	trainable_linear_quant	Trainable linear quantization (TLQ).
unary_analysis	unary_analysis	Unary (non-quantized) sensitivity analysis.

4. Save Formats (spec.save[] Dispatched by type)

type	Configuration	Applicable Protocol
ascendv1_saver	ascendv1_saver	All protocols (modelslim_v1, multimodal_*, modelslim_convert).
compressed_tensors	compressed_tensors	Only modelslim_v1 and modelslim_convert; the VLM / multimodal generation protocols do not accept it.
mindie_format_saver	mindie_format_saver	All protocols.

5. Auto Tuning

For the top-level protocol and field overview, see Auto Tuning Configuration Reference Navigation. The following are field references for each strategy and evaluation service:

Configuration	Description
tuning_plan	Tuning plan root configuration (strategy + evaluation).
strategy_standing_high	The standing_high strategy.
strategy_standing_high_with_experience	The standing_high_with_experience strategy.
strategy_binary_fallback	The binary_fallback strategy.
evaluation_service_oriented	AISBench evaluation service and pre-check configuration.