Skip to content

v35_v40_implementation.md

Latest commit

 

History

History
79 lines (59 loc) · 4.86 KB

v35_v40_implementation.md

File metadata and controls

79 lines (59 loc) · 4.86 KB
Raw

v35.0.0–v40.0.0 Implementation Record

This document records the completed development scope for MLX Server Manager v35.0.0 through v40.0.0. The app version in the development tree is 40.0.0. The latest published binary remains v34.0.0 until a separate commit, tag, push, packaging, and release workflow is explicitly requested.

v35.0.0 — Capability-Aware Profile Compatibility

  • Generalized runtime option probing across mlx_lm.server and Rapid-MLX.
  • Compares the optional flags requested by a profile with the flags exposed by the configured executable.
  • Treats unknown support as unknown rather than unsupported.
  • Rejects stale compatibility evidence when the executable path changes.
  • Blocks start only for known unsupported requested options.
  • Provides an explicit unsaved-draft cleanup action; saving remains a separate user action.

v36.0.0 — Structured Launch-Failure Diagnosis

  • Captures a bounded in-memory window of managed-process output during launch.
  • Classifies common executable, dependency, model-file, port, permission, parser, memory, readiness, and early-exit failures.
  • Redacts bearer tokens and secret-like values from copied diagnostic output.
  • Connects diagnosis to failed start, failed readiness, and unexpected managed-process termination paths.
  • Does not persist raw command output or send diagnostics externally.

v37.0.0 — Registered Model Health

  • Adds an explicit health check for registered profiles only.
  • Inspects configured local paths without scanning unrelated folders.
  • Reports present, missing, incomplete, duplicate-path, and external-identifier states.
  • Reuses the existing local-model integrity boundary for required configuration, tokenizer, and weight files.
  • Does not download, delete, relocate, or repair model files.

v38.0.0 — Start Resource Forecast

  • Reuses existing multi-model memory assessment and KV-cache safety logic.
  • Shows projected managed memory, available-memory headroom, and conservative cache guidance before start.
  • Can place safer values or a Memory Saver template into an unsaved profile draft.
  • Never silently changes the saved profile or bypasses existing critical-memory confirmation.

v39.0.0 — Benchmark-Guided Profile Tuning

  • Produces tuning suggestions only from current, comparable benchmark sessions with at least two successful samples.
  • Excludes stale execution fingerprints, changed hardware/backend conditions, thermally pressured runs, and insufficient samples.
  • Maps supported use-case evidence to existing explicit profile templates.
  • Applies recommendations only to an unsaved draft for review.

v40.0.0 — Operational Readiness

  • Adds one explicit, non-generating readiness surface covering runtime readiness, profile compatibility, registered-model health, endpoint state, resource forecast, prior-session recovery candidates, benchmark evidence, latest launch diagnosis, and Hermes Agent configuration.
  • Runs independent readiness, compatibility, and model-health work concurrently behind a replaceable coordinator.
  • Discards stale completion when settings, profiles, or selected profile change during the run.
  • Does not start a server, send a generation request, download a model, adopt an endpoint, stop a process, or save settings.
  • Keeps an already-running selected managed endpoint from being misclassified as blocked merely because its port is occupied.

Persistence

  • Advances the authoritative settings snapshot to schema 2.
  • Reads and migrates schema 1 snapshots to schema 2.
  • Retains authoritative-state and recovery-backup behavior.
  • Rejects unsupported future schemas rather than overwriting them.

Safety Boundaries

  • Direct Mode remains unchanged.
  • Process termination remains limited to app-owned managed processes.
  • External and recovered endpoints never gain process ownership.
  • Capability cleanup, resource guidance, and benchmark tuning modify unsaved drafts only.
  • Model health is explicit and limited to registered paths.
  • Launch diagnosis remains bounded, memory-only, copy-redacted, and local.
  • No telemetry, external log sending, inference proxy, background generation, silent downloads, or automatic persistence was added.

Verification Completed

  • Full Debug test suite succeeded.
  • Focused v35–v40 tests succeeded, including executable-path staleness handling.
  • Release build succeeded with code signing disabled for local verification.
  • Static analysis succeeded.
  • Strict Concurrency Debug build succeeded with Swift and GCC warnings treated as errors.
  • English/Japanese localization guardrails and rendering tests succeeded as part of the full suite.
  • git diff --check succeeded.

Real-model launch behavior, model-specific memory estimates, and benchmark recommendations still depend on the user's installed runtimes, registered model files, and local Mac conditions. Those environment-specific checks are not represented as universal compatibility claims.