This is the first public release of the redesigned ML Inference Advisor core package.
Highlights
- Split MLIA into a core package and independently maintained target, backend, and converter plugins.
- Added runtime plugin discovery and post-analysis plugin support.
- Introduced a public Python API for analysis and capability discovery.
- Migrated the CLI to Typer with consolidated target and backend command groups.
- Added extensible support for measured profiling data.
- Added TOML, environment-variable, and .env configuration.
- Added versioned documentation and an embedded CycloneDX SBOM.
Standardized output
MLIA Core now treats standardized output as the authoritative analysis result. This release adds:
- Structured result-level advice.
- Explicit metric availability.
- Standard inference-latency and model-weight-memory metrics.
- Canonical source-operator identities for differnet model formats.
- Entity provenance, relationships, stack traces, and stricter validation.
- Configurable entity collapse and metric-breakdown projection.
Improvements and fixes
- Improved CLI completion and target discovery performance.
- Added typed, dynamically registered backend options.
- Added automatic backend installation support.
- Improved terminal encoding, output width, NO_COLOR, and redirected-output handling.
- Fixed logging-resource cleanup between runs.
- Improved backend archive handling.
- Improved standardized-output projection correctness.
- Added release-artifact validation and security scanning.
Compatibility
MLIA requires Python 3.10 or newer. Target-specific functionality is no longer included in the core package and is provided by separate plugins.
See the README (README.md) for installation and an overview of the plugin ecosystem. Detailed CLI, API, architecture, configuration, and output guidance is available in the documentation (docs/README.md).