mlx-swift-lm

Fix LFM2-MoE sigmoid routing and expert-bias selection (#332)

Align Falcon H1 with upstream MLX inference (#331)

fix Gemma4 prepare() to honor windowSize -- prompt prefill is never chunked (#337)

Add Gemma 4 MTP speculative decoding (#308)

Fix Gemma 4 cross-layer KV sharing in no-cache forward pass (#333)