Training engine
Can I fine-tune this LLM?
Training needs gradients and optimizer state as well as weights. Compare the estimated memory for full fine-tuning, LoRA and QLoRA on your hardware.
Training setup
Memory estimates compare methods; a fit does not guarantee framework support on every device.
Llama · Llama 3.1 Community
Hardware picker is loading. Please wait.
Pick your hardware above to see which fine-tuning methods fit.
Running and training need different memory
Fine-tuning holds gradients and optimizer state, often several times the weights themselves. A model you can run can still be out of reach to fully fine-tune. The three method estimates use the same batch, sequence length, rank, optimizer and checkpointing settings so you can compare their memory requirements.
Just want to run it? Back to the inference calculator →