LoRA Training

Resume LoRA training from a checkpoint

Continue a previous LoRA training run from its last saved checkpoint — for extending training after hitting a max-cost limit, or fine-tuning further after review.

Estimated $1–$6 per resumed run · 10–30 min

What inputs you need

  • Checkpoint file from a prior run
  • Same or updated dataset
  • Training config (steps to add)

What Badgr returns

  • Updated LoRA safetensors file
  • Combined training loss curve
  • Sample generations at new checkpoints

Recommended GPU routes

Estimated cost: $1–$6 per resumed run · Estimated runtime: 10–30 min

Example command

badgr run "python train_lora.py --resume checkpoint.safetensors --config lora.yaml" \
  --gpu A100_40GB \
  --max-cost 6 \
  --max-runtime 40

Common setup failures Badgr avoids

Optimizer state mismatch breaking a resumed run — Badgr's training image saves optimizer state alongside weights so resume behaves like an uninterrupted run.

Learning-rate schedule restarting from step zero on resume — config carries forward the original step count so the schedule continues correctly.

Ready to run this?

Launch from the dashboard, CLI, or Compute API. Max-cost protection included.