Tool
VRAM calculator
How much GPU memory does fine-tuning need?
Roughly 2.2 GB per billion parameters for half-precision LoRA, 1.7 GB for four-bit, and about 16 GB for full fine-tuning. Preference tuning roughly doubles whichever applies, because a frozen reference model is held alongside. Pick a model below for the exact admission thresholds.
Half precision
—
—
Four-bit
—
—
Classes that will run it
| Class | VRAM | Half precision | Four-bit | Training / hr |
|---|
If nothing fits
- Reduce sequence length to what the data actually needs. Setting it to the model maximum by default is the most common source of waste — check the token-length percentiles in the validation report first.
- Turn on gradient checkpointing. It discards activations and recomputes them, cutting the largest variable term for roughly a third more time.
- Use gradient accumulation. Several small batches summed into one update give the same effective batch at a fraction of the peak memory.
Frequently asked questions
Why does preference tuning need twice the memory?
It holds a frozen reference copy of the model alongside the one being trained, to measure and penalise drift. Two models resident means roughly twice the memory, which is why a model that fits supervised training on a mainstream card often needs the next class up for DPO.
What if nothing fits?
Three levers, in this order: reduce sequence length to what the data actually needs, turn on gradient checkpointing, and use gradient accumulation. Together they will usually bring a run back inside a class you can afford.
Are these numbers guaranteed?
They are admission thresholds — the point below which a run will not be scheduled because it would fail. They are deliberately conservative, and they do not include the attention cache at serving time, which grows with context length and concurrency.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.