Fine-tuning & training
Fine-tune — LoRA, QLoRA, DPO, Axolotl, Unsloth
Open-source fine-tuning that just runs, with the right CUDA on the right GPU.
Launch single-gpu lora / qlora (7b–13b)
Pay by the hour · no contracts · terminate anytime from the dashboard
What this is
Fine-tuning a 7B–70B model is where dependency hell hurts most — a Flash Attention mismatch can cost you a whole evening. The instance boots on a stock PyTorch 2.2 + CUDA 12.1 base — Axolotl, Unsloth, and DeepSpeed aren't on disk yet, but the combination below is the known-good pairing, so your first commands are a pip install, then the dataset and the yaml.
Our value-add
We publish a pinned PyTorch + CUDA + Axolotl/DeepSpeed combination that's known to work together, so you're not the one debugging library skew.
Environment & recommended stack
- PyTorch 2.2 + CUDA 12.1 (stock base image)boots clean — the fine-tuning stack below is a pip install away
- Axolotl + DeepSpeed + Flash Attention 2recommended — full LoRA/QLoRA/DPO matrix
- Unslothrecommended for fastest single-GPU 7B fine-tune
- bitsandbytes + PEFT + TRLrecommended — the standard LoRA/QLoRA stack
- wandb CLIpip install wandb, then paste your API key
You stay in control
- Bring your own dataset — HF hub, S3, or scp into the instance
- Multi-GPU training — the 8x H100 NVLink box runs DeepSpeed Zero-3
- Auto-destroy deadline — set it to 20% above your expected training time
- Prepaid credits are a hard fail-safe — you get a low-balance warning email, and the instance is destroyed at $0 instead of overdrawing you
Example workflow
Kick off an Axolotl LoRA run
# on the instance after ssh:
pip install "axolotl[deepspeed]"
cd ~ && git clone https://huggingface.co/datasets/your-team/sft-data
cat > config.yaml <<'YAML'
base_model: meta-llama/Llama-3.1-8B
load_in_4bit: true
adapter: qlora
datasets:
- path: ~/sft-data
YAML
accelerate launch -m axolotl.cli.train config.yamlRun inside the instance after `ssh root@<instance-ip>` or via the web terminal.
Launch options
Pick a GPU for fine-tuning & training
Recommended start
Single-GPU LoRA / QLoRA (7B–13B)
One H100 80 GB comfortably QLoRA-tunes a 13B at good throughput
$4.41/ hour
Launch with this configScale option
Multi-GPU full fine-tune / 70B
8× H100 + NVLink for Zero-3 full-parameter or 70B LoRA
$31.10/ hour
Launch with this configHourly rates shown are current target rates — the authoritative price is surfaced on the /launch page at the moment you provision. Billing is per 5 minutes of runtime.