jundot/omlx v0.1.4

oMLX v0.1.4

on GitHub

What's New

Anthropic API

• Add thinking block support for Anthropic /v1/messages endpoint

Sampling and generation

• Add repetition_penalty support with automatic generation_config.json loader
• Fix top_k not being passed through to SamplingParams in BatchedEngine (#20)

CORS

• Wire cors_origins config into FastAPI CORSMiddleware (#18)

Port conflict handling

• Detect port conflicts on startup and offer adopt/kill options
• Fix killing the correct process on conflict and improve status indicator color

SSD cache

• Fix SSD cache load_block deadlock with async I/O and Metal-safe writes (#16)

macOS app

• Add memory compatibility badges to recommended models
• Allow clearing per-model settings back to default values

Tests

• Add tests for repetition_penalty, port conflict detection, and sampling params

New Contributors

• @thornad made their first contribution in #20
• @rndmcnlly made their first contribution in #18