5 min read

Which GPU to pick for LoRA training: SDXL, FLUX and 7B models

Choosing a card for LoRA training on SDXL, FLUX and 7B LLMs: memory requirements, QLoRA versus LoRA, and how to estimate training time before you start.

LoRA trains a small add-on to a frozen model, so it needs far less memory than full training. Choosing a card comes down to three things: base model size, precision and batch size.

Memory guide

SDXL LoRA: 16 GB or more.
FLUX LoRA: 24 GB or more; less is possible with offloading, but slower.
QLoRA for a 7B model at 4 bits: 16 GB or more.
LoRA for a 7B model at fp16: 24 GB or more.
Full training of a 7B model: about 112 GB, so several GPUs.

Estimating time before you start

Run 20–50 steps on your dataset with your batch size.
Measure the seconds per step.
Multiply by the total step count and add 20% for saving and validation.

That tells you the cost before you commit to the full run. For short training, a slightly more expensive but faster card often comes out cheaper overall.