Full fine-tuning
Fine-tuning that updates every weight in the model, not a small adapter. Highest quality ceiling, but needs several times the model size in VRAM and usually a multi-GPU node.
What it is
Every parameter is trainable. In addition to the weights you must hold gradients and optimiser state (with Adam, ~2x the weights again), so peak VRAM is roughly 4x the model size in fp16 before activation memory. A 70B model needs ~8x 80 GB GPUs; techniques like ZeRO / FSDP shard this across the node.
Why it matters
Full fine-tuning is worth it when a lora adapter plateaus below your quality bar - deep behavioural change, a new language, a very different output distribution. For most business tasks LoRA/QLoRA get you there for a tenth of the cost.