用 LoRA/QLoRA 微调大模型，并借助思维链数据生成与蒸馏提升推理准确率，适合低成本定制模型。
单卡微调大模型前的显存预检工具，估算显存占用并给出可直接跑的 LoRA/QLoRA 配置，避免盲目开训就爆显存。
面向 Apple Silicon 的本地大模型推理与微调框架，用 Rust 实现，支持 LoRA/QLoRA 微调、量化、模型服务和 MLX/Metal 加速。
…recommended CPU full/LoRA, and RTX 4080 CUDA LoRA/QLoRA. Training cases completed two optimizer updates, changed saved tensors, and reloaded for generation; QLoRA exercis
…ed on Linux and Windows. Local RTX 4080 CUDA LoRA/QLoRA candidate checks execute updates, save and reload outside the checkout. WSL GPU, other GPU hardware, Flash Attenti
…hout rescaling static weights - fail clearly when QLoRA or 8-bit optimizer dependencies are unavailable - add safer recipe generation and optional Liger kernels - expand
…nting does not remove it. Previously unmodeled. - QLoRA weights are no longer "all params × 0.5 B" — only Linear layers are 4-bit; embeddings, lm head (twice for untied m
…s-foundation/bitsandbytes/pull/1856 [Docs Update] QLoRA 4-bit Support on ROCm by @Abdennacer-Badaoui in https://github.com/bitsandbytes-foundation/bitsandbytes/pull/1857
…n be costly. Maestro's built-in support for LoRA, QLoRA, and graph freezing allows training larger models even on less powerful hardware. - VLMs fine-tuning requires lots