unary_analysis Configuration Reference¶

1. Configuration Overview¶

Unary (non-quantized) sensitivity analysis processor configuration.

Item	Content
Config Class	`UnaryAnalysisProcessorConfig`
Source	processor.py

Field Path	Type	Required/Optional	Default	Value Range or Format	Description	Referenced Config
`type`	`string`	Optional	`unary_analysis`	`unary_analysis`	Processor type, fixed to `unary_analysis`.	None
`metrics`	`string`	Optional	`kurtosis`	—	Analysis metric: `quantile` (quantile), `std` (standard deviation), `kurtosis` (kurtosis), `ra_compress` (RA Compress long-sequence compression head screening)	None
`patterns`	`list[string]`	Optional	`['*']`	—	List of layer name patterns to analyze; the default `*` matches all	None
`metric_params`	`object`	Optional	`{}`	—	Metric-specific hyperparameters, parsed only by the analysis method specified in `metrics`. Currently supported only for `metrics=ra_compress`: `induction_head_ratio` (default 0.14) and `echo_head_ratio` (default 0.01), both with a value range of [0, 1].	None

Configuration Constraints

Validates metric-specific hyperparameters; illegal use raises an error directly.

apiversion: modelslim_v1
spec:
  process:
  - type: unary_analysis
    metrics: kurtosis
    patterns:
    - '*'
    metric_params: {}

metric_params takes effect only for metrics=ra_compress; passing it with other metrics raises an error directly. This field is set only in the configuration, and the CLI does not expose a corresponding parameter; msmodelslim analyze attn_head runs with the defaults (induction_head_ratio=0.14, echo_head_ratio=0.01).
A larger ratio selects (keeps) more KV heads, which means lower KV Cache compression strength, smaller accuracy risk, and smaller memory savings; a smaller ratio does the opposite. It is recommended to evaluate empirically based on the number of selected heads and the downstream compression ratio.
The command-line workflow reads the template msmodelslim/core/analysis_service/pipeline_analysis/pipeline_template/ra_compress.yaml installed with the package; after modifying this YAML, you must re-run bash install.sh in the repository root for the change to take effect.

ra_compress overrides the default ratios:

process:
- type: unary_analysis
  metrics: ra_compress
  metric_params:
    induction_head_ratio: 0.14
    echo_head_ratio: 0.01