import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_compute_dtype=torch.bfloat16)
tok = AutoTokenizer.from_pretrained("pieroot/timesorter-qwen3.5-4b-sft-v4")
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", quantization_config=bnb,
                                            device_map="auto", dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, "pieroot/timesorter-qwen3.5-4b-sft-v4").eval()

Qwen3.5는 추론 시 enable_thinking=False (chat template) 권장 — 깨끗한 JSON 출력.

학습 코드·데이터셋: https://github.com/jung-geun/TimeSorter