Skip to content

fix: Prepare certified inactive embedding targets - #822

Draft
buildingjoshbetter wants to merge 1 commit into
fix/model-transfer-retention-publish-297-v1from
fix/prepared-target-publish-795-v1
Draft

buildingjoshbetter wants to merge 1 commit into
fix/model-transfer-retention-publish-297-v1from
fix/prepared-target-publish-795-v1

Conversation

@buildingjoshbetter

Copy link
Copy Markdown
Owner

Tier rebuilds need to prepare an inactive embedding target without changing the serving tier or silently accepting a custom-model fallback. This prerequisite adds an immutable target descriptor, explicit daemon certification with actual output-width checks, and an encode-only local lease. Named completion/separation tables initialize independently of active configuration, with paired rollback and case-insensitive TEMP-shadow rejection.

The daemon reuses its existing cache, admission, deadline and CPU-recovery controls. Local preparation permits one outstanding lease and shares exact-model ownership with existing inference. The optional timed ownership path preserves legacy callers' locking behavior. Existing models, rerankers, dimensions and retrieval policy are unchanged.

Stacked on #820. Related to #795; this does not close it. No tier-switch caller adopts this API yet. Full-corpus retention, truthful activation and the original Mac memory incident remain unresolved. No whole-process byte bound or native MPS improvement is claimed.

Validation:

  • Sequential offline GPUBox installed-wheel builds and full suites passed. Publication: 2,285 tests and 696 subtests; combined tree: 3,109 tests and 2,671 subtests. Both had 17 skips and 3 deselections, real sqlite-vec enabled, no timeout and no OOM kill.
  • Independent review passed 37 saved probes plus two legacy lock-protocol checks. The new suite includes actual sqlite-vec pair initialization, idempotence, wrong-width refusal and owned/borrowed rollback.
  • Original-source negative control: 27 assertion failures, zero errors. Removed width, permission, close and TEMP-case guards also fail their targeted assertions. Six existing fixtures remain byte-identical, preserving 809 assertions.
  • The first combined full run failed five unchanged ownership tests after default lock acquisition changed. Restoring the original untimed path fixed them; failed evidence and assertions are preserved. Earlier review failures are also retained.
  • Ruff, Python 3.10 parsing, exact publication mapping, source-hash, privacy and public commit-identity checks passed.

Full log SHA256: publication 993a79e7c930fb6b8604e7ddd306238e84ab59f020f3e43862e4eeaad7ed5b0d; combined 4a6586ba85124e2190f7ce4b16f621caf8bbd8d9e21b357685af9f67d8b84cba.

Draft only. Nothing is merged or released.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant