MLX-LM – v0.30.1
🔥 MLX LM v0.30.1 is LIVE — Apple Silicon LLMs just got a massive upgrade!
🚀 New Models: RWKV7, Solar Open, K-EXAONE MoE, IQuest Coder V1, YoutuLLM + Minimax M2 (perfect for long-context chats)!
💬 Chat Fixes: Custom DSV32 templates work, non-standard tokenizers behave, and `generation_config` errors? Ignored — no more crashes.
⚡ Performance: GIL starvation fixed in `_generate`, Phi3 (LongRoPE) batched prompts now stable, and `load_config` checks files first — smarter loading.
✨ New Features: `logits_processors` in `batch_generate` (fine-tune outputs like a pro), `model-path` flag for cleaner conversions, and support for mxfp8 & nvfp4 quantization — squeeze more power from your M-series chip!
📋 Bug fixes: `/v1/models` now shows local models correctly.
Big thanks to new contributors: @cubist38, @vyaivanov, @sjugin, @jaycoolslm, @MollySophia, @cxl-git-hub, @lazarust!
Upgrade. Tinker. Crush your next LLM project. 🛠️💻