model_name_or_path: Qwen/Qwen3-4B-Instruct-2507
adapter_name_or_path: saves/qwen3-4b/lora/sft
template: qwen3_nothink
infer_backend: huggingface  # choices: [huggingface, vllm, sglang, ktransformers]
trust_remote_code: true
