MLX-LM-LoRA v5.8.6: Apple Silicon fine-tuning library jumps from 3.1.3, adds DSLA preference alignment and critic-free KLPO RL

MLX-LM-LoRA is an Apple MLX library for training LLMs on Apple Silicon. v5.8.6 (Oct 5, 2026) is the first release after v3.1.3 and focuses on post-training. New: DSLA, which adds prompt-response similarity and batch-direction latent supervision on top of a DPO, ORPO or CPO objective (DSLA-DPO uses a frozen reference; ORPO/CPO are reference-free), and KLPO, a critic-free RL method with token- or sequence-level regression and MC, binary, top-k or full KL estimators, with no GRPO group normalization, PPO clipping or reference model. GRPO scoring is more stable; Online DPO, XPO, RLHF REINFORCE and PPO gain microbatching; gated-delta layers get a fast VJP with a checkpointed fallback. QAT now covers DSLA, a new docs site includes a Python API reference, and the license is Apache 2.0. No speed or quality benchmarks were published. Upgrade with pip install -U mlx-lm-lora; it needs mlx>=0.32.3, mlx_lm>=0.32.0 and Python>=3.11.