#!/usr/bin/env python3
"""Compare one selected head and one control across several synthetic sequences."""
import argparse
import hashlib
import json
from pathlib import Path
import torch
import transformers
from transformers import AutoModelForCausalLM
import induction_heads as induction
import induction_prefix as prefix


def main():
    p=argparse.ArgumentParser(description=__doc__)
    p.add_argument('--discover-seed',type=int,default=7)
    p.add_argument('--test-seeds',type=int,nargs='+',default=[17,31,41,53,67])
    p.add_argument('--control',type=int,nargs=2,default=[0,0],metavar=('LAG','HEAD'))
    p.add_argument('--output',type=Path,default=Path('induction-prefix-sweep.json'))
    p.add_argument('--offline',action='store_true')
    a=p.parse_args()
    if a.output.exists():p.error('Outputfilen findes allerede. Vælg et andet navn.')
    if not 3<=len(a.test_seeds)<=10 or len(set(a.test_seeds))!=len(a.test_seeds) or a.discover_seed in a.test_seeds or any(not 0<=x<2**32 for x in [a.discover_seed,*a.test_seeds]):
        p.error('Vælg 3–10 forskellige testseeds, forskellige fra udvælgelsesseed, mellem 0 og 2³²−1.')
    if not 0<=a.control[0]<6 or not 0<=a.control[1]<8:p.error('Kontrolhead skal være lag 0–5 og head 0–7.')
    torch.set_num_threads(1)
    model=AutoModelForCausalLM.from_pretrained(induction.MODEL,revision=induction.REVISION,local_files_only=a.offline,attn_implementation='eager').eval()
    try:results=[prefix.run(model,a.discover_seed,seed,a.control) for seed in a.test_seeds]
    except ValueError as e:p.error(str(e))
    result={'format':'learning-lab-induction-prefix-sweep-v1','sweep_sha256':hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),'results':results}
    with a.output.open('x',encoding='utf-8') as f:json.dump(result,f,ensure_ascii=False,indent=2,allow_nan=False);f.write('\n')
    print(f"{len(results)} testfølger; valgt head {results[0]['selected']}; gemt {a.output}")
if __name__=='__main__':main()
