> BROMANDER_LABS
CAN YOUR RIGTRAIN IT?
Fine-tuning eats far more memory than inference — weights, gradients, optimizer state, and activations all have to live on the card at once. Here's the honest peak VRAM for full, LoRA, and QLoRA across 50+ models.
Math runs in your browserMixed precision · fp32 master copy
01 — Model & method
02 — Batch & sequence
03 — Your hardware
Peak VRAM
QLoRA · peak training memory
6.03 GB
Trainable
8M
of total
0.10%
GPU usable
22.1 GB
Fits on one NVIDIA RTX 4090. 6.03 GB of 22.1 GB usable VRAM.
Memory breakdown
Weights · 67%Gradients · 0%Optimizer · 2%Activations · 11%Overhead · 21%
Weights
4.03 GB
Gradients
0.02 GB
Optimizer + master
0.10 GB
Activations
0.64 GB
Overhead
1.24 GB
Estimate assumes bf16 mixed precision with an fp32 master weight copy. Activation memory is approximate — it varies with attention implementation (Flash Attention lowers it) and exact model architecture.
─── Your Training Card ───
The actual image that shows on X, LinkedIn, and Facebook when you share the link.
Live preview
Need a training pipeline built?
Talk to Bromander Studios