Skip to content

v0.5.0 — Speech-to-Text (mlx-whisper) & Real Image Generation (mflux)

Choose a tag to compare

View all tags
@vdparikh vdparikh released this 04 Sep 18:12
· 135 commits to main since this release
v0.5.0
7c38726

Track 3: Real Modalities on Apple Silicon Metal

1. Local Speech-to-Text (/v1/audio/transcriptions)

  • Implemented OpenAI-compatible POST /v1/audio/transcriptions endpoint (multipart/form-data) supporting json, verbose_json, text, vtt, and srt formats.
  • Added MLXWhisperRuntime for real on-device dictation via mlx-whisper with Metal GPU acceleration.
  • Added EchoAudioTranscriptionRuntime for fast, zero-weight offline tests.
  • Curated vdplabs.whisper-tiny.compact.v1 and registered aliases whisper-1, whisper-compact, and transcribe-compact.
  • Added pantry transcribe <audio-file> CLI command.

2. Real Image Generation (/v1/images/generations)

  • Implemented MFluxImageRuntime powering real FLUX.1 (schnell/dev) and SD-Turbo image generation on Apple Silicon Metal with 8-bit/4-bit quantization.
  • Retained deterministic EchoImageRuntime for zero-weight smoke testing.
  • Curated vdplabs.flux1-schnell.standard.v1 with aliases flux-schnell, image-standard, and flux.
  • Added pantry image <prompt> CLI command.

3. Core Dependencies & Homebrew Packaging

  • Added python-multipart to dependencies for robust multipart upload handling in FastAPI.
  • Added optional extras [whisper] and [mflux].
  • Updated Homebrew tap formula pantry.rb.
Assets 2
Loading