modelslim_v1 Configuration Reference¶

1. Configuration Overview¶

modelslim_v1 quantization task configuration, located at the YAML root.

Item	Content
Config Class	`ModelslimV1QuantConfig`
Source	quant_config.py

modelslim_v1 is the new-generation quantization processing framework introduced by the quantization tool and is currently evolving quickly. Compared with modelslim_v0, modelslim_v1 has the following advantages:

Algorithms are implemented independently, and configurations can be combined freely.
Layer-wise quantization is supported, greatly reducing resource consumption.
It does not depend on a specific version of CANN.

2. Parameter List¶

2.1 ModelslimV1QuantConfig

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
`apiversion`	`string`	Optional	`modelslim_v1`	`modelslim_v1`	—	None
`spec`	`object`	Required	None	—	The spec structure of the `modelslim_v1` service; declares the quantization pipeline, save format, and calibration data.	This page §2.2

Configuration Constraints

None.

2.2 ModelslimV1ServiceConfig

The spec structure of the modelslim_v1 service; declares the quantization pipeline, save format, and calibration data.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
`runner`	`string`	Optional	`auto`	`auto`, `model_wise`, `layer_wise`, `dp_layer_wise`	Pipeline execution mode: `auto` selects automatically based on the device count (single device `layer_wise`, multiple devices `dp_layer_wise`), `model_wise` computes the whole model at once, `layer_wise` computes layer by layer, and `dp_layer_wise` computes layer by layer with data parallelism.	None
`prior`	`list[object]`	Optional	`[]`	—	List of prior stages; each stage contains process and dataset	This page §2.2.1
`process`	`list[object]`	Optional	`[]`	—	Quantization processor chain, executed in order; each element is a processor configuration dispatched by `type`, such as `linear_quant`, `awq`, `smooth_quant`, and more.	This page §2.2.2
`save`	`list[object]`	Optional	`[]`	—	Save format list; each element is a save format configuration dispatched by `type`, such as `ascendv1_saver`, `compressed_tensors`, `mindie_format_saver`.	This page §2.2.3
`dataset`	`string`	Optional	`mix_calib.jsonl`	—	Calibration dataset name (the file name under `lab_calib`) or a dataset path, used for parameter estimation and sensitivity calibration during quantization.	None

Configuration Constraints

None.

2.2.1 PriorStageConfig

Prior-stage configuration: process + dataset only, used for prior stages such as adapt_rotation stage1.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
`process`	`list[object]`	Optional	`[]`	—	List of processors for this stage	This page §2.2.2
`dataset`	`string / null`	Optional	`null`	—	Dataset name for this stage; if not provided, spec.dataset is used	None

Configuration Constraints

None.

2.2.2 AutoProcessorConfig

Derived Classes

Config Class	`type`	Description	Document
`AdaptRotationProcessorConfig`	`adapt_rotation`	Adaptive rotation (adapt_rotation) processor configuration.	adapt_rotation Configuration Reference
`AutoroundProcessorConfig`	`autoround_quant`	AutoRound quantization processor configuration.	autoround_quant Configuration Reference
`AWQProcessorConfig`	`awq`	AWQ (Activation-aware Weight Quantization) processor configuration.	awq Configuration Reference
`BinaryAnalysisProcessorConfig`	`binary_analysis`	Binary (quantized/non-quantized) sensitivity analysis processor configuration.	binary_analysis Configuration Reference
`BinaryOperatorLayerWiseProcessorConfig`	`binary_operator_layer_wise`	Layer-wise sensitivity analysis processor configuration (compares per-block floating-point and quantized outputs).	binary_operator_layer_wise Configuration Reference
`BinaryOperatorModelWiseProcessorConfig`	`binary_operator_model_wise`	Model-level sensitivity analysis configuration (compares final model outputs, using the MSE metric)	binary_operator_model_wise Configuration Reference
`DynamicCacheProcessorConfig`	`dynamic_cache`	KV cache quantization processor configuration.	dynamic_cache Configuration Reference
`FA3QuantProcessorConfig`	`fa3_quant`	FA3 (FlashAttention-3) quantization processor configuration.	fa3_quant Configuration Reference
`FlatQuantProcessorConfig`	`flatquant`	FlatQuant processor configuration: defines quantization training parameters, strategy, mixed precision, and more	flatquant Configuration Reference
`FlexAWQSSZProcessorConfig`	`flex_awq_ssz`	Flex AWQ SSZ outlier suppression processor configuration.	flex_awq_ssz Configuration Reference
`FlexSmoothQuantProcessorConfig`	`flex_smooth_quant`	FlexSmoothQuant smooth quantization processor configuration.	flex_smooth_quant Configuration Reference
`FloatSparseProcessorConfig`	`float_sparse`	Floating-point sparsity processor configuration.	float_sparse Configuration Reference
`GroupProcessorConfig`	`group`	Processor combiner configuration.	group Configuration Reference
`IterSmoothProcessorConfig`	`iter_smooth`	Iterative smooth (IterativeSmooth) processor configuration.	iter_smooth Configuration Reference
`KVSmoothProcessorConfig`	`kv_smooth`	KV cache smoothing processor configuration.	kv_smooth Configuration Reference
`LinearProcessorConfig`	`linear_quant`	Linear layer quantization processor configuration.	linear_quant Configuration Reference
`LoadProcessorConfig`	`load`	Module load/unload processor configuration.	load Configuration Reference
`OASQProcessorConfig`	`oasq`	OASQ (Outlier-Aware Smooth Quantization) processor configuration.	oasq Configuration Reference
`OnlineQuaRotProcessorConfig`	`online_quarot`	Online QuaRot rotation processor configuration.	online_quarot Configuration Reference
`QuaRotProcessorConfig`	`quarot`	QuaRot (offline rotation) processor configuration.	quarot Configuration Reference
`QuantSaveProcessorConfig`	`saver`	Unified save processor configuration.	saver Configuration Reference
`SmoothQuantProcessorConfig`	`smooth_quant`	SmoothQuant smooth quantization processor configuration.	smooth_quant Configuration Reference
`SVDResidualProcessorConfig`	`svd_res`	SVD residual (low-rank compensation) processor configuration.	svd_res Configuration Reference
`TrainableLinearQuantProcessorConfig`	`trainable_linear_quant`	Trainable linear quantization (TLQ) processor configuration.	trainable_linear_quant Configuration Reference
`UnaryAnalysisProcessorConfig`	`unary_analysis`	Unary (non-quantized) sensitivity analysis processor configuration.	unary_analysis Configuration Reference

2.2.3 QuantFormatConfig

Derived Classes

Config Class	`type`	Description	Document
`AscendV1QuantFormatConfig`	`ascendv1_saver`	AscendV1 save format configuration; exports weight files in the Ascend on-disk format.	ascendv1_saver Configuration Reference
`CompressedTensorsQuantFormatConfig`	`compressed_tensors`	compressed_tensors save format configuration; exports safetensors weights and config.json.	compressed_tensors Configuration Reference
`MindIEQuantFormatConfig`	`mindie_format_saver`	MindIE save format configuration; exports weight files in the MindIE on-disk format.	mindie_format_saver Configuration Reference

3. Complete Configuration Reference¶

apiversion: modelslim_v1
spec:
  runner: auto
  prior: []
  process: []
  save: []
  dataset: mix_calib.jsonl