adapt_rotation Configuration Reference

1. Configuration Overview

Adaptive rotation (adapt_rotation) processor configuration.

Item	Content
Config Class	AdaptRotationProcessorConfig
Source	adapt_rotation.py

2. Parameter List

2.1 AdaptRotationProcessorConfig

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
type	string	Optional	adapt_rotation	adapt_rotation	Processor type, fixed to adapt_rotation.	None
stage	int	Required	None	1, 2	Rotation adaptation stage: 1 or 2; determines which stage configuration is used.	None
stage_config	object	Required	None	—	Stage configuration object (an internally assembled field; it is required but generated automatically by a before-validator based on stage, so users do not configure it in YAML); do not configure this field directly in YAML. Instead, flatten the stage fields (such as steps and quant_dtype) under the processor; see AdaptRotationStage1ProcessorConfig Configuration Reference / AdaptRotationStage2ProcessorConfig Configuration Reference.	This page §2.2

Configuration Constraints

• Assembles the flat fields into stage_config by stage: only fields of the corresponding stage are allowed; extra fields raise an error.

2.2 ProcessorConfig

Derived Classes

Config Class	Description	Document
AdaptRotationStage1ProcessorConfig	Configuration of adapt_rotation stage 1.	This page §2.2.1
AdaptRotationStage2ProcessorConfig	Configuration of adapt_rotation stage 2.	This page §2.2.2

2.2.1 AdaptRotationStage1ProcessorConfig

Configuration of adapt_rotation stage 1.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
type	string	Optional	_adapt_rotation_stage1	_adapt_rotation_stage1	Stage configuration type; an internal identifier that does not need to be configured.	None
steps	int	Optional	20	—	Number of iterative optimization steps	None
quant_dtype	string	Optional	int4	int4, int8	Quantization bit width; it should match the activation quantization type used downstream (for example int4 for w4a4 and int8 for w8a8)	None
layer_type	list[string]	Optional	['up_proj']	At least 1 item	List of layer name substrings whose activations are collected	None
block_size	int	Optional	-1	—	Block size; -1 means hidden_dim	None
max_samples	int	Optional	2048	—	Maximum number of samples per layer	None

Configuration Constraints

• Validates layer_type: each element must be a non-empty string of length <= 128
• Validates block_size: the value must be -1 or a non-negative integer power of 2

2.2.2 AdaptRotationStage2ProcessorConfig

Configuration of adapt_rotation stage 2.

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
type	string	Optional	_adapt_rotation_stage2	_adapt_rotation_stage2	Stage configuration type; an internal identifier that does not need to be configured.	None
online	bool	Optional	false	—	Whether to enable online rotation	None
block_size	int	Optional	-1	—	Block size; -1 means hidden_dim	None
down_proj_online_layers	list[int]	Optional	[]	—	Index list of online down_proj layers	None
max_tp_size	int	Optional	4	—	Maximum TP parallelism	None

Configuration Constraints

• Validates down_proj_online_layers: each element must be a non-negative integer
• Validates max_tp_size: it must be at least 1 and a power of 2
• Validates block_size: the value must be -1 or a non-negative integer power of 2

3. Complete Configuration Reference

apiversion: modelslim_v1
spec:
  prior:
  - process:
    - type: adapt_rotation
      stage: 1
      steps: 20
      quant_dtype: int4
      layer_type:
      - up_proj
      block_size: -1
      max_samples: 2048
  process:
  - type: adapt_rotation
    stage: 2
    online: false
    block_size: -1
    max_tp_size: 1