Anassbzdd has released FitCheck, a tool that estimates peak VRAM requirements for LoRA, QLoRA, and full fine-tuning of large language models without requiring a GPU or PyTorch. The estimator analyzes the model's Hugging Face config.json and parameter-count metadata to report memory breakdowns, usable capacity, headroom, and the largest estimated micro-batch that fits.

  • Supports serving estimates based on model weights and KV cache.
  • Includes an advisor to explore batch size, sequence length, and LoRA rank.
  • Achieved a 2.4% mean absolute error across 57 calibration runs against real Tesla T4 measurements.
  • Produced 12/12 correct fit-boundary verdicts in a separate 12-run holdout test.

The tool helps users determine if their configuration fits in their GPU's memory before starting an LLM job, aiming to reduce guessing and out-of-memory errors.