modelslim_convert Configuration Reference

1. Configuration Overview

modelslim_convert quantization (weight conversion) task configuration, located at the YAML root.

Item	Content
Config Class	ModelslimConvertQuantConfig
Source	quant_config.py

2. Parameter List

2.1 ModelslimConvertQuantConfig

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
apiversion	string	Optional	modelslim_convert	modelslim_convert	Weight conversion task type, fixed to modelslim_convert.	None
spec	object	Required	None	—	The spec structure of the modelslim_convert service. Declares weight name renaming/transformation (preprocess), linear layer conversion rules (linears), the save format (save), parallel execution (parallel), and defaults (defaults).	This page §2.2

Configuration Constraints

• None.

2.2 ModelslimConvertServiceConfig

The spec structure of the modelslim_convert service.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
preprocess	list[object]	Optional	[]	—	List of preprocessing steps; each item has type rename or convert.	This page §2.2.1
linears	list[object]	Optional	[]	—	List of linear layer conversion rules.	This page §2.2.6
save	list[object]	Optional	[]	—	List of save format configurations; the first one takes effect.	This page §2.2.7
parallel	object	Optional	See the nested configuration default	—	CPU parallelism configuration. For NPU multi-card, use the command line --device_id.	This page §2.2.8
defaults	object	Optional	See the nested configuration default	—	Global default values used when fields are omitted.	This page §2.2.9

Configuration Constraints

• None.

2.2.1 PreprocessConfig

Derived Classes

Config Class	type	Description	Document
RenamePreprocessConfig	rename	One of the modelslim_convert preprocessing steps: batch-renames weight tensors.	This page §2.2.2
ConvertPreprocessConfig	convert	One of the modelslim_convert preprocessing steps: applies weight transformations (split/merge, and so on) to matching linear layers.	This page §2.2.4

2.2.2 RenamePreprocessConfig

One of the modelslim_convert preprocessing steps: batch-renames weight tensors.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
type	string	Optional	rename	rename	Preprocessing type, fixed to rename.	None
patterns	list[object]	Optional	[]	—	List of renaming rules, applied one by one to matching weight names.	This page §2.2.3

Configuration Constraints

• None.

2.2.3 RenamePattern

Weight tensor renaming rule: rewrites weight names matching from to to.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
from	string	Required	None	—	Source weight name pattern; wildcards are supported. Matching weight names are rewritten.	None
to	string	Required	None	—	Target weight name pattern after rewriting.	None

Configuration Constraints

• None.

2.2.4 ConvertPreprocessConfig

One of the modelslim_convert preprocessing steps: applies weight transformations (split/merge, and so on) to matching linear layers.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
type	string	Optional	convert	convert	Preprocessing type, fixed to convert.	None
source	list[string]	Optional	[]	—	List of source weight name patterns (the linear layers to transform).	None
target	list[string]	Optional	[]	—	List of target weight name patterns (the transformation result).	None
ops	list[object]	Optional	[]	—	List of weight transformation operators, such as chunk and merge.	This page §2.2.5

Configuration Constraints

• None.

2.2.5 ConvertOpConfig

Weight operators in the convert preprocessing step, such as splitting/merging fused gate/up projections.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
type	string	Required	None	—	Operator type: chunk (splits fused gate/up), merge (merges gate/up), or another mapping operator.	None
dim	int / null	Optional	null	—	Split/merge dimension: when unspecified it is inferred from the operator type, being 1 for chunk and 0 for merge.	None
projections	list[string] / null	Optional	null	—	Names of the projections produced by chunk: when unspecified they are inferred as gate_proj and up_proj.	None

Configuration Constraints

• None.

2.2.6 LinearConvertConfig

Specifies the rule for converting matching linear layers to the target IR.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
match	list[string]	Optional	[]	—	List of matching linear layer name patterns.	None
target	string	Required	None	FLOAT, FP8_BLOCK, W8A8_MXFP8, W4A4_MXFP4, W4A8_MXFP8, INT4_PACKED, NVFP4_MODELOPT, HIFP4, UNKNOWN	Target IR type for conversion, such as W8A8_MXFP8 or INT4_PACKED.	None
route	list[string] / string	Optional	auto	FLOAT, FP8_BLOCK, W8A8_MXFP8, W4A4_MXFP4, W4A8_MXFP8, INT4_PACKED, NVFP4_MODELOPT, HIFP4, UNKNOWN; auto	Conversion route: an explicit IR list (the first element is the source IR), or auto to let the virtual tree infer it from the weight dtype.	None

Configuration Constraints

• None.

2.2.7 SaveConfig

Save format configuration of modelslim_convert.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
type	string	Optional	ascend_v1	—	Save format: ascend_v1 (Ascend; equivalent to ascendv1 in ConvertDefaults.dst_format); compressed_tensors (HF-compatible safetensors); huggingface/hf are aliases of compressed_tensors.	None
part_file_size	int	Optional	4	—	Shard file size in GB; 0 means no sharding.	None

Configuration Constraints

• None.

2.2.8 ParallelSpecConfig

Parallel execution configuration of modelslim_convert. The device (NPU/CPU and card index) is determined by the CLI --device / --device_id; the YAML contains no device field.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
cpu_workers	int	Optional	8	—	Number of parallel workers for CPU conversion: 1 means a single process; greater than 1 means multiple processes across groups plus threads within a group. Takes effect only with --device cpu; NPU conversion always uses one subprocess per card (determined by --device_id) and does not need this setting.	None
max_group_size	int / null	Optional	null	—	Maximum number of tasks in a single dependency group; beyond this the group is split into several sub-groups spread across processes. Unset means no splitting.	None

Configuration Constraints

• cpu_workers takes effect only with --device cpu. For NPU use --device npu --device_id ....
• max_group_size: for models with many expert tasks, such as MoE, set it to roughly "expert tasks per layer / number of cards" (for example 96 for 8 cards) to avoid long-tail waiting on a single process/card for a large group.

2.2.9 ConvertDefaults

Global default values used when conversion rules do not explicitly declare fields.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
src_format	string	Optional	auto	auto	Source weight format: auto infers it from the model adapter or the weight directory.	None
dst_format	string	Optional	ascendv1	ascendv1, ascendv1_saver, compressed_tensors, huggingface, hf	Target save format: ascendv1 (Ascend; equivalent to ascend_v1 in SaveConfig.type); compressed_tensors (HF-compatible safetensors); huggingface/hf are aliases of compressed_tensors.	None
dst_ir	string / null	Optional	null	FLOAT, FP8_BLOCK, W8A8_MXFP8, W4A4_MXFP4, W4A8_MXFP8, INT4_PACKED, NVFP4_MODELOPT, HIFP4, UNKNOWN, or null	Target IR type; when unset it is determined by the target format.	None

Configuration Constraints

• None.

3. Complete Configuration Reference

apiversion: modelslim_convert
spec:
  linears:
  - match:
    - model.language_model.layers.*.mlp.gate_proj
    - model.language_model.layers.*.mlp.up_proj
    - model.language_model.layers.*.mlp.down_proj
    - model.language_model.layers.*.mlp.shared_experts.gate_proj
    - model.language_model.layers.*.mlp.shared_experts.up_proj
    - model.language_model.layers.*.mlp.shared_experts.down_proj
    - model.language_model.layers.*.mlp.experts.*.gate_proj
    - model.language_model.layers.*.mlp.experts.*.up_proj
    - model.language_model.layers.*.mlp.experts.*.down_proj
    - model.language_model.layers.*.self_attn.q_a_proj
    - model.language_model.layers.*.self_attn.q_b_proj
    - model.language_model.layers.*.self_attn.kv_a_proj_with_mqa
    target: W8A8_MXFP8
    route: auto
  save:
  - type: ascend_v1
    part_file_size: 4
  parallel:
    cpu_workers: 8
    max_group_size: null
  defaults:
    src_format: auto
    dst_format: ascendv1
    dst_ir: null