llm-gpu-lab
An end-to-end LLM toolkit that runs entirely on one consumer NVIDIA GPU, taking a user from environment diagnosis through pretraining, fine-tuning, evaluation, GGUF export and a self-contained benchmark report.
llm-gpu-lab walks a developer with a single consumer GPU, in the 12 to 24 gigabyte class such as an RTX 4080, 4070 Ti Super or 3090, through a full local LLM workflow with no paid hosted inference and no cloud training: environment diagnosis, tokenizer training, a tiny GPT pretrained from scratch, generation from that checkpoint, LoRA or QLoRA supervised fine-tuning of a small open model, lightweight evaluation with optional lm-evaluation-harness integration, GGUF export, and llama.cpp serving, finishing with GPU benchmarking and a reproducible HTML report.
Two parallel paths, a from-scratch TinyGPT pretraining path with its own trained tokenizer and an open-model LoRA or QLoRA fine-tuning path using each base model's own Hugging Face tokenizer, share the same benchmark and reporting infrastructure, with an optional GGUF export and llama.cpp serving branch off the fine-tuning path. Every pipeline step writes a JSON artifact, and every number in the final report is read from one of those artifacts rather than a hand-entered or expected value; a smoke pipeline covering the fine-tuning through evaluation steps runs in about 30 seconds of compute plus a one-time download of a small model.