The three choices in a fine-tune on a rented GPU are the method, the tool and where the result lives afterwards. This section has one guide for a language model, one for an image model, and the rules that apply to both on DeployGPU.
Pick the method by memory: QLoRA, LoRA or full
Full fine-tuning updates every weight. Hugging Face's memory guide counts about 18 bytes per parameter for it before activations, so an 8 billion parameter model needs 144 GB (calculated), more than one 80 GB card holds. LoRA freezes the base model and trains small adapter matrices, so it costs the base in 16-bit plus a margin. QLoRA holds the base in 4-bit, and its paper reports fine-tuning a 65B model on a single 48 GB GPU. Unsloth's requirements page lists 22 GB for QLoRA on a 27B model against 64 GB for 16-bit LoRA, and all three sources were checked on 1 October 2026. Start with QLoRA, which Unsloth's guide also recommends, and move up only when the results fall short. The VRAM page maps every figure to the cards DeployGPU launches.
Pick the tool: Unsloth, Axolotl, LLaMA-Factory or TRL
Four open-source projects fine-tune text models. Unsloth is a framework for running and training models, and its guide here runs QLoRA from install to adapter. Axolotl is a fine-tuning framework driven by one YAML file, with full fine-tuning, LoRA and QLoRA among its methods. LLaMA-Factory fine-tunes 100+ models from a command line or a web UI. TRL is Hugging Face's library of post-training trainers such as SFTTrainer, DPOTrainer and GRPOTrainer. Image models use a different toolchain, and the image LoRA guide covers AI Toolkit and kohya_ss for FLUX and SDXL.
What a DeployGPU VM limits: 1, 2, 4 or 8 GPUs
A VM has 1, 2, 4 or 8 GPUs, and the smallest class has 48 GB. There is no multi-node training and no self-serve InfiniBand, so a job that needs more than 8 GPUs or a fixed term goes through the quote form on the clusters page. Full fine-tuning of a larger model runs on one multi-GPU VM with FSDP or DeepSpeed, as the multi-GPU guide shows.
Keep the result: stopping deletes the disk
Stopping an instance terminates it and deletes its disk, including adapters, checkpoints and datasets. Push adapters to the Hugging Face Hub or copy them to your own machine before you stop, and save checkpoints off the VM during long runs, as checkpoint and resume explains. The stop guide has the full checklist.