LitGPT — обучение и файн-тюнинг языковых моделей

★ 7.7 · ai-tooling

litgpt is a Claude Code skill that implements and trains LLMs using Lightning AI's LitGPT, covering 20+ pretrained architectures including Llama, Gemma, Phi, Qwen, and Mistral. All implementations are single-file with no abstraction layers, making the transformer internals readable and suitable for educational exploration. The skill covers the full workflow: downloading models via `litgpt download`, full fine-tuning with `litgpt finetune`, memory-efficient LoRA/QLoRA adaptation (`litgpt finetune_lora`) with configurable rank from 8 to 64 and optional weight merging, and pretraining from scratch on custom corpora with multi-GPU FSDP support. Dependencies are litgpt, torch, and transformers. It suits ML engineers and researchers who need reproducible LLM experiments, production fine-tuning on a single 8 GB+ GPU, or clean reference code for understanding how large language models work.