Skip to content

ci.md

Latest commit

 

History

History
64 lines (51 loc) · 3.02 KB

ci.md

File metadata and controls

64 lines (51 loc) · 3.02 KB

Continuous Integration & Validation

VECTORIA uses GitHub Actions not just for build hygiene, but as a formal Correctness Proof mechanism.

Architecture-Aware Testing

CI is split into architecture-specific workflows to reflect the reality of hardware availability.

ARM64 NEON (Validated)

  • Runner: macos-latest (Apple Silicon)
  • Scope: Full SIMD validation.
  • Status: Results in this environment are considered Proven.

x86_64 AVX2 (Validated)

  • Runner: ubuntu-latest
  • Scope: Full SIMD validation (detected at runtime).
  • Status: Results in this environment are considered Proven.

Tests Executed on CI

Every commit triggers the following suites on both validated architectures:

  • test_gemm_simd: Compares SIMD kernels against bit-exact Reference.
  • test_multi_op: Verifies composite graph execution and deterministic scheduling.
  • test_graph_equivalence: Ensures Reference/SIMD parity within validated tolerance.
  • test_determinism_stress: Asserts bitwise reproducibility over 50 iterations.
  • High-Level Semantic Suites (Reference-Only):
    • test_layernorm: Statistical property and broadcast validation.
    • test_attention: Numerical equivalence for Scaled Dot-Product expansion.
    • test_multi_head_attention: Structural splitting and join validation.
    • test_transformer_encoder: Full block integration proof.

Tooling Validation

Standalone tooling is validated on every commit to ensure that introspection capabilities remain consistent with the frozen semantic surface.

  • tooling: Runs schema validation, trace analysis, and determinism diffing on simulated execution traces.

Packaging Validation

The distribution pipeline is validated to ensure that prebuilt binaries and Python/Swift packages are correctly formed and installable.

  • packaging: Builds and installs the Python wheel, verifying the vectoria-trace CLI functionality.

Supported Validation Platforms

CI explicitly proves correctness on:

  • macOS (Latest): ARM64 NEON backend + CoreML lowering.
  • Ubuntu (Latest): x86_64 AVX2 backend.

Reproducibility Checklist

To achieve bitwise identity between a local environment and CI:

  1. Fixed Binary: Use the exact same compiled artifact.
  2. Identical Architecture: Compare ARM64 to ARM64 or x86_64 to x86_64.
  3. Kernel Policy: Ensure both use KernelPolicy::Reference if cross-architecture identity is required.
  4. No OS Noise: VECTORIA's single-threaded model eliminates OS scheduling noise from the result.

Scope of CI Validation

Tested:

  • Bit-exact Reference correctness.
  • SIMD vs. Reference parity within tolerance.
  • Structural op metadata preservation.
  • CoreML MIL generation and validation.

NOT Tested:

  • Multi-threaded execution (Engine is strictly serial).
  • Third-party library interference (NumPy/PyTorch internal math).
  • Driver-level optimization overrides.

Validation Logs

Logs from correctness tests are preserved as CI artifacts. These logs are part of the scientific provenance of every VECTORIA release.