Unified runner

from safetune.steer import run

texts = run(
    wrapped_model,
    prompts=["How to make a bomb?"],
    backend="hf",              # "hf", "vllm-hook", "vllm-logits"
    tokenizer=tokenizer,
    max_new_tokens=256,
)

backend	What it does
"hf"	Plain transformers generation
"vllm-hook"	Activation steering in vLLM
"vllm-logits"	Decoding steering in vLLM

For backend="vllm-hook", the SteerSpec is auto-extracted from the wrapped model. For backend="vllm-logits", pass a DecodeSteerSpec explicitly via decode_spec= (it is not auto-extracted). Both vLLM backends also need the base model_id=.