Swift Fine-tuning (ms-swift)

Swift provides concise scripts to fine-tune models with LoRA/QLoRA and export weights for serving.

Install

pip install ms-swift  # or refer to repo instructions

Minimal LoRA Run

swift finetune \
  --model "meta-llama/Llama-3-8b" \
  --dataset path/to/train.jsonl \
  --output ./outputs/llama3-lora \
  --batch-size 4 --accum 8 --epochs 1 \
  --lr 2e-4 --lora-r 8 --lora-alpha 32 --bf16

Export and Serve

swift export --input ./outputs/llama3-lora --merge-lora --out ./merged
# Then load with vLLM/TGI or via LiteLLM routing.

Notes

• Use QLoRA for low VRAM; validate outputs with small eval sets.
• Track dataset and hyperparams in the output folder for reproducibility.