"""Fresh-process control replay: distinguish visual hallucination from conversation carryover."""
from pathlib import Path
import sys
import time

PROJECT=Path(__file__).resolve().parents[1];sys.path.insert(0,str(PROJECT))
from pallet.mask_model_training_data import read,write,jsonl,require,gpu_receipt
from pallet.cluster_targets import digest
from pallet.pallet_vlm_native_intensity import ROOT,parse_structure


def main():
    import torch
    from PIL import Image
    from transformers import AutoProcessor,AutoModelForImageTextToText
    out=ROOT/'cold_control';require(not out.exists(),'Preserve cold replay')
    torch.set_num_threads(4);torch.manual_seed(20261105);gpu=gpu_receipt('m2')
    out.mkdir(mode=0o700)
    folder=ROOT/'models/qwen4b';pinned=read(folder.parent/'qwen4b_download.json')
    require(all(digest(folder/n)==h for n,h in pinned['files'].items()),'Pretrained model changed')
    tasks={t['task_id']:t for t in jsonl(ROOT/'sample/tasks.jsonl')}
    order=['S001_blank_structure','S003_blank_structure','S001_structure','S001_blank_structure','S001_blank_crop']
    write(out/'plan.json',dict(order=order,fresh_process=True,new_conversation_each_request=True,no_prompt_change=True))
    processor=AutoProcessor.from_pretrained(folder,local_files_only=True);processor.image_processor.min_pixels=128*128;processor.image_processor.max_pixels=1024*1024
    model=AutoModelForImageTextToText.from_pretrained(folder,dtype=torch.bfloat16,attn_implementation='sdpa',local_files_only=True).cuda().eval();model.requires_grad_(False)
    results=[];start=time.monotonic()
    with torch.inference_mode():
        for i,key in enumerate(order):
            t=tasks[key];image=Image.open(ROOT/t['image']).convert('RGB')
            messages=[dict(role='user',content=[dict(type='image',image=image),dict(type='text',text=t['prompt'])])]
            text=processor.apply_chat_template(messages,tokenize=False,add_generation_prompt=True,enable_thinking=False)
            data=processor(text=[text],images=[image],return_tensors='pt').to('cuda')
            ids=model.generate(**data,max_new_tokens=t['max_new_tokens'],do_sample=False,pad_token_id=processor.tokenizer.pad_token_id)
            raw=processor.tokenizer.decode(ids[0,data['input_ids'].shape[1]:],skip_special_tokens=True)
            original=read(ROOT/'runs/qwen4b/predictions'/(key+'.json'))
            value=dict(sequence_index=i,task_id=key,prompt=t['prompt'],input_image_sha256=t['image_sha256'],raw=raw,exact_main_raw_match=raw==original['raw'])
            if t['kind']=='structure':value['parsed']=parse_structure(raw)
            results.append(value)
    write(out/'predictions.jsonl',results,lines=True)
    write(out/'complete.json',dict(state='fresh_process_control_replayed',model='qwen4b',gpu=gpu,tasks=len(results),
        all_main_outputs_exactly_reproduced=all(p['exact_main_raw_match'] for p in results),
        first_blank_pallet=results[0]['parsed']['fields']['托盘'],same_blank_before_after_real=results[0]['raw']==results[3]['raw'],
        seconds=time.monotonic()-start,predictions_sha256=digest(out/'predictions.jsonl'),code_sha256=digest(__file__)))
    print(read(out/'complete.json'),flush=True)


if __name__=='__main__':main()
