~/runthismodel
daemon okbuild 5a3c91d00:00:00Z

Fine-tuning

Fine-tune VRAM calculator

Inference VRAM is the easy half — fine-tuning eats 5–20× more. We model the four real cost lines (weights, gradients, optimizer state, activations) for QLoRA, LoRA, and full fine-tunes.

Setup

Method

VRAM needed

6.1 GB

Smallest GPU class that fits: RTX 4060 Ti 16 GB

  • Weights3.73 GB
  • Gradients0.07 GB
  • Optimizer state0.15 GB
  • Activations1.18 GB

Tips to reduce VRAM

  • QLoRA is 8–10× lighter than full fine-tune for ~95 % of the quality
  • Gradient checkpointing trades ~30 % VRAM for ~25 % slower steps
  • Reduce batch size to 1 + use gradient accumulation
  • AdamW 8-bit (bitsandbytes) cuts optimizer state 4×
  • Shorter sequences during training, then ramp up

Estimates ±20 %. Real numbers depend on framework (Unsloth saves 30–60 %), gradient checkpointing, and CPU-offload settings. Inference VRAM for this model → · Rent a GPU big enough →