Fine-tune with Axolotl

Use swm to find the cheapest GPU, provision a pod, and run a fine-tuning job with Axolotl.

Steps

1. Find the cheapest A100

swm gpus -g a100 --sort price

2. Create a pod

swm pod create -p <cheapest-provider> -g "A100 80GB" -n finetune \
  --lifecycle auto-down --idle-timeout 60 -y

3. Install Axolotl

swm setup install axolotl <provider>:<pod-id>

4. Upload your config and data

swm upload <provider>:<pod-id> ./config.yml /workspace/
swm upload <provider>:<pod-id> ./dataset/ /workspace/dataset/ -r

5. Run training

swm run <provider>:<pod-id> "cd /workspace && axolotl train config.yml"

6. Save and terminate

swm pod down finetune

Your checkpoints are pushed to S3. Resume on a different GPU anytime.