Skip to content

versioning.md

Latest commit

 

History

History
60 lines (45 loc) · 2.46 KB

versioning.md

File metadata and controls

60 lines (45 loc) · 2.46 KB

MIND Core Stability & Versioning

MIND Core follows the mind-spec Core v1 contract (mind-spec/spec/v1.0). This page documents how stability is applied for the compiler/runtime pipeline and how versions are communicated in the CLI.

Semantic Versioning (0.y.z)

MIND Core currently publishes 0.y.z versions with the following rules:

  • Patch (0.y.*): bug fixes, documentation updates, and tooling-only changes that do not alter IR semantics or CLI contracts.
  • Minor (0.y.*): additive IR extensions are allowed, including new instructions and flags. Existing IR semantics must remain compatible.

What counts as breaking

  • Changing the semantics of an existing IR instruction, its shape rules, or its broadcasting/reduction behavior.
  • Altering CLI output formats (textual IR, diagnostics, or stable flags) in ways that break downstream tooling.
  • Modifying MLIR lowering patterns that are relied on by snapshot tests or downstream compiler passes.

Stability categories

Stable surfaces

  • Core IR v1: instruction set, shape rules, reductions, broadcasting, and verification guarantees.
  • Autodiff API: differentiate_function, GradientResult, and AutodiffError.
  • Canonicalization guarantees: deterministic rewrites prior to lowering.
  • mindc base flags and textual IR output.
  • Compiler diagnostics: error codes (E1xxx…E5xxx) and the JSON diagnostic field names for the --diagnostic-format json surface.
  • Core v1 GPU profile: DeviceKind/BackendTarget enum variants for CPU and GPU, the runtime backend-selection error model (e.g., BackendUnavailable), and the GPUBackend trait surface that executes canonical IR on GPU devices.
  • Conformance suite: the golden Core v1 conformance corpus shipped with this repository. Passing the suite for a given profile (CPU baseline or the GPU profile) demonstrates API, IR, autodiff, and MLIR stability for that profile in the current release.

Conditionally stable surfaces

  • MLIR lowering (mlir-lowering feature): stable within a minor version but may change across minor releases to track backend needs.

Experimental surfaces

  • New operations added to the IR.
  • New feature flags exposed by the CLI or libraries.
  • Concrete GPU or accelerator backend implementations and device-specific lowering pipelines.

References