Fine-tune open-weight models on your own GPU pod

How to fine-tune a model on Aquanode

1. 1

   Size the GPU for your model and method

   LoRA/QLoRA fits on far less VRAM than a full fine-tune of the same model. Use the GPU recommender to see native/FP8/INT4 VRAM for a specific model, or the training cost calculator to estimate a full run.

2. 2

   Spin up a pod

   Start from an empty pod (a GPU and a shell) or from the notebook environment (JupyterLab with PyTorch preinstalled) if you'd rather work interactively. Either way you get root access and per-second billing.

3. 3

   Install your fine-tuning stack

   Aquanode doesn't preinstall a fine-tuning framework. On your pod, pip/conda install whatever your recipe needs: Hugging Face Transformers and PEFT for LoRA/QLoRA, TRL for SFT/DPO, or a higher-level trainer like Axolotl or Unsloth.

4. 4

   Train, then checkpoint

   Run your training job on the pod. Snapshot the workspace with workspace backups so a checkpoint survives past the run, or to move it to a different GPU.

5. 5

   Run the result

   Load your fine-tuned weights back into an inference stack on a pod sized for them, the same way you'd run any other model.

Models people fine-tune