LoRA and QLoRA
Parameter-efficient methods train small adapters on top of a frozen model. A 7–13B model fine-tunes on one RTX PRO 6000 or one H100; 70B with QLoRA fits on a single 80–96 GB card.
Full fine-tuning
Updating every weight needs memory for gradients and optimizer state, roughly 16 bytes per parameter. A 7B model needs 2–4 H100s; 70B needs a full 8-GPU H100 or H200 node with sharding.
Typical run lengths
Most fine-tunes take hours to a few days. Rent on demand, billed per minute, and stop the server when the run finishes.





