core-labs/core
core-labs/core
1
0
Fork 0
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions
core/docs/benchmarks/apple-uma-mlx-exact-recall-local-validation.md
Shay b30716a19c
feat(bench): add MLX exact CGA recall experiment (#909) 
* feat(bench): add MLX exact CGA recall experiment

* test(bench): cover MLX exact recall experiment contracts

* docs(bench): document MLX exact recall experiment

* docs(bench): record MLX experiment integration status

* docs(bench): add MLX local validation handoff
2026-06-24 13:53:50 -07:00

1.2 KiB
Raw Blame History

Gemini Local Validation Handoff — MLX Exact Recall

Branch: feat/apple-mlx-exact-cga-recall-experiment

This branch is deliberately narrow. It adds the MLX exact CGA recall experiment as an isolated benchmark module before wiring it into the main Apple UMA report.

Commands

cd /Users/kaizenpro/Projects/core
git fetch origin --prune
git switch feat/apple-mlx-exact-cga-recall-experiment
git pull --ff-only origin feat/apple-mlx-exact-cga-recall-experiment

uv run python -m pytest -q tests/test_apple_uma_mlx_exact_recall.py
uv run python -m benchmarks.apple_uma_mlx_exact_recall --json

If MLX is available or installable locally:

CORE_BACKEND=rust uv run python -m benchmarks.apple_uma_mlx_exact_recall --json

Touch-up limits

Only patch:

  • MLX import/API mismatches
  • formatting or lint issues
  • test failure fixes directly tied to this module

Do not add serving integration. Do not add Metal. Do not add CoreML/ANE. Do not change semantic source of truth. Do not use ANN or approximate recall.

Success criteria

  • Skip path passes when MLX is unavailable.
  • MLX-present path emits cases.
  • Every emitted case has parity.parity_pass: true.
  • Copy-in/copy-out fields remain explicit.
  • Non-claims remain visible.