github jundot/omlx v0.2.13

latest releases: v0.3.6, v0.3.5, v0.3.5-rc1...
one month ago

Download the DMG that matches your macOS version (sequoia or tahoe).
If you're on an M5 Mac, you must use the macos26-tahoe DMG for M5 Neural Accelerator.

Highlights: real-time prefill tracking and VLM chat support

  • Real-time prefill progress tracking in the dashboard. See all loaded models, per-request prefill progress bars with speed and ETA, memory usage, and active/waiting request counts at a glance.

  • Image upload support in Chat with oMLX. VLM models are auto-detected and you can attach multiple images per message. Thanks to @Leslie-SSS for the contribution! (#226)

New Features

Active Models card in Dashboard (#233)

  • New card showing all loaded models with status (idle, active, prefilling, loading)
  • Per-model memory size and unload button on hover
  • Real-time per-request prefill progress bars with tok/s speed and ETA
  • Memory usage bar showing model memory budget utilization
  • Active/waiting request count footer
  • Prefill tracking uses cpu counters only (via mlx-lm's prompt_progress_callback), zero impact on inference performance

VLM image upload in Chat (#226)

  • Auto-detect VLM models and show image upload button
  • Multi-image support with content array format
  • Image-only requests without text
  • Prevent localStorage quota exceeded when saving chat with images

Dashboard UI refresh

  • Status tab renamed to Dashboard
  • Benchmark table headers now have tooltip info icons (#242)
  • Chat theme system refactored from hardcoded Tailwind colors to CSS variables

Bug Fixes

  • fix cache-layer count validation for hybrid models (was using total layers instead of cache layers)
  • fix MiniCPM-o position state not initialized for text-only requests, causing shape errors (#241)
  • fix unsupported TTS/ASR models (e.g. Dia, Kokoro) crashing model discovery (#239)
  • fix models directory getting wiped on "reset all settings"
  • fix repetitive admin stats access logs flooding the console
  • fix Tailwind CSS build not scanning JS files, causing dynamic classes (e.g. text-3xl) to be missing
  • fix noisy TRACE log for models without custom settings

New Contributors

Thanks to @Leslie-SSS for the contribution!

full changelog: v0.2.12...v0.2.13

Don't miss a new omlx release

NewReleases is sending notifications on new releases.