from types import SimpleNamespace
import json

import torch

from pallet.pallet_description_vlm_repair import label_mask


def tokenizer(text,**kwargs):return dict(offset_mapping=[(i,i+1) for i in range(len(text))])


def data(answer):
    prompt={'input_ids':torch.tensor([[1,2]])}
    full={'input_ids':torch.tensor([[1,2]+[ord(c) for c in answer]+[3]])}
    return full,prompt


def test_missing_value_only_mask_keeps_quotes_braces_and_EOS():
    answer=json.dumps({'结构':'三支腿木质托盘','观察':None},ensure_ascii=False,separators=(',',':'));full,prompt=data(answer)
    labels=label_mask(SimpleNamespace(tokenizer=tokenizer),full,prompt,answer,False)
    masked=''.join(chr(v) for v in full['input_ids'][0,2:][labels[0,2:]==-100]);supervised=labels[0][labels[0]!=-100].tolist()
    assert masked=='null' and supervised[-1]==3 and ord('}') in supervised and ord('"') in supervised
    assert (labels[0,:2]==-100).all()


def test_real_observation_all_answer_tokens_are_supervised():
    answer='{"结构":"托盘","观察":"叉齿遮挡"}';full,prompt=data(answer)
    labels=label_mask(SimpleNamespace(tokenizer=tokenizer),full,prompt,answer,True)
    assert torch.equal(labels[:,2:],full['input_ids'][:,2:]) and (labels[:,:2]==-100).all()
