跳转至

Auto Tuning Configuration Reference Navigation

This directory is the field-level reference for msModelSlim auto-tuning configuration (YAML). It is organized as: write the tuning plan first, then look up strategies and evaluation services by type. For the usage workflow and end-to-end steps, see Auto Tuning Usage Guide; for command-line parameters, see msmodelslim tune Command-Line API Reference; for strategy selection, see Auto Tuning Strategy Overview. For the fields of quantization task YAML, see Quantization Task Configuration Reference Navigation; for the master tables of processors and save formats, see Configuration Reference Navigation.

1. How to Use This Documentation Set

  1. Determine the top-level structure: the root of a tuning YAML is not apiversion but the two required fields of tuning_plan: strategy and evaluation.
  2. Look up strategies: strategy is dispatched by type. Use the tuning strategies table below to find the corresponding page.
  3. Look up evaluation services: evaluation is dispatched by type. Use the evaluation services table below to find the corresponding page. Currently only service_oriented is available.
  4. Look up embedded quantization configuration: the template of standing_high and binary_fallback references a best-practice or modelslim_v1 task configuration; for protocol pages, see Quantization Task Configuration Reference Navigation; for processor type values, see Configuration Reference Navigation. standing_high_with_experience does not require a complete hand-written quantization YAML; it is generated from quant_type and structure_configs.

Except for this page, every page in this directory is 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. Tuning Plan

Configuration Description
tuning_plan Tuning plan root configuration, containing only strategy and evaluation.

3. Tuning Strategies (strategy Dispatched by type)

type Configuration Purpose
standing_high standing_high Standing high: run sensitive layer analysis first, then roll back progressively and try different outlier suppression chains; a quantization template must be provided.
standing_high_with_experience standing_high_with_experience Experience-based Standing high: generates a quantization plan from the quantization type and structure configuration, with no need to hand-write a complete template.
binary_fallback binary_fallback Binary fallback: searches for the shortest fallback prefix over a complete PracticeConfig template.

standing_high.template points to modelslim_v1 Configuration Reference. binary_fallback.template points to PracticeConfig Configuration Reference, and apiversion must be modelslim_v1.

4. Evaluation Services (evaluation Dispatched by type)

type Configuration Purpose
service_oriented evaluation_service_oriented Service-oriented evaluation: accuracy expectations, AISBench evaluation, and the vLLM-Ascend inference engine.