Skip to content

1.1.0

Latest

Choose a tag to compare

@NerdSnipe NerdSnipe released this 07 Aug 06:05

Highlights

  • Dual residency slots (primary/auxiliary) with per-slot wired-memory reservation tickets, bounding process-wide wired memory to the device's recommended working set
  • Pre-check wired-memory admission so a denied reservation can never hang or leak
  • Explicit MLXProvider.releaseResidency(_:) API to force-release a residency slot
  • MLXProvider.residentWeightBytes(in:) and MLXProvider.modelResidency APIs
  • adapterDirectoryURL on MLXProvider with auto-load on first inference, graceful degradation on missing/corrupt adapters
  • FunctionGemma stream processor, native tool-call schema/round-tripping fixes
  • Bound Memory.memoryLimit/cacheLimit to the device's actual recommended working set

Fixes

  • Fix duplicate AIChatKit identity in monorepo builds
  • Fix SPI manifest authors field to valid string format