#!/usr/bin/env python3
"""Map residual-stream patch effects across token positions and layers."""
import argparse
import csv
import hashlib
import json
from datetime import datetime, timezone
from pathlib import Path

import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
import pythia_intervention as patch


def prepare(tokenizer, source_text, target_text, source_answer, target_answer):
    source=tokenizer(source_text,return_tensors='pt',add_special_tokens=False)
    target=tokenizer(target_text,return_tensors='pt',add_special_tokens=False)
    a_ids=source['input_ids'][0].tolist();b_ids=target['input_ids'][0].tolist()
    if len(a_ids)!=len(b_ids) or not 2<=len(a_ids)<=16:raise ValueError('Prompts skal have samme tokenlængde mellem 2 og 16.')
    differences=[i for i,(a,b) in enumerate(zip(a_ids,b_ids)) if a!=b]
    if len(differences)!=1 or differences[0]==len(a_ids)-1:raise ValueError('Prompts skal afvige i præcis ét token, som ikke er det sidste.')
    a_answer,b_answer=patch.validate_answer_boundaries(tokenizer,source_text,target_text,a_ids,b_ids,source_answer,target_answer)
    return source,target,a_ids,b_ids,differences[0],a_answer,b_answer


def measure(model, tokenizer, source_text, target_text, source_answer, target_answer):
    source,target,a_ids,b_ids,changed,a_answer,b_answer=prepare(tokenizer,source_text,target_text,source_answer,target_answer)
    clean,cache_a=patch.capture(model,source);baseline,cache_b=patch.capture(model,target)
    if len(cache_a)!=6 or len(cache_b)!=6:raise ValueError('Forventede seks lag i checkpointet.')
    if not torch.isfinite(clean).all() or not torch.isfinite(baseline).all():raise ValueError('Baseline-logits er ikke endelige.')
    def diff(logits):return float(logits[a_answer]-logits[b_answer])
    base=diff(baseline);rows=[];final_source_error=None
    for layer in range(6):
        for position in range(len(a_ids)):
            patched=patch.intervene(model,target,layer,position,cache_a[layer]);identity=patch.intervene(model,target,layer,position,cache_b[layer])
            if not torch.isfinite(patched).all() or not torch.isfinite(identity).all():raise ValueError('Et indgreb gav ikke-endelige logits.')
            if layer==5 and position==len(a_ids)-1:final_source_error=patch.identity_logit_max(patched,clean)
            value=diff(patched);rows.append({'layer':layer,'position':position,'patched_difference':value,'effect':value-base,'identity_logit_max':patch.identity_logit_max(identity,baseline)})
    return {'format':'learning-lab-pythia-position-map-v1','model':patch.MODEL,'revision':patch.REVISION,'torch_version':str(torch.__version__),'transformers_version':transformers.__version__,
            'source_text':source_text,'target_text':target_text,'source_answer':source_answer,'target_answer':target_answer,
            'source_token_ids':a_ids,'target_token_ids':b_ids,'source_tokens':[tokenizer.decode([i]) for i in a_ids],'target_tokens':[tokenizer.decode([i]) for i in b_ids],
            'changed_position':changed,'boundary_checked':True,'source_answer_id':a_answer,'target_answer_id':b_answer,'source_difference':diff(clean),'target_difference':base,'source_top_token':tokenizer.decode([int(clean.argmax())]),'target_top_token':tokenizer.decode([int(baseline.argmax())]),'final_source_logit_max':final_source_error,'rows':rows}


def main():
    p=argparse.ArgumentParser(description='Kortlæg patching i seks lag og alle tokenpositioner.')
    p.add_argument('--offline',action='store_true');p.add_argument('--source');p.add_argument('--target');p.add_argument('--source-answer');p.add_argument('--target-answer');p.add_argument('--case',type=Path,help='Brug en casefil fra tokenkontrollen');p.add_argument('--output-dir',type=Path);args=p.parse_args()
    if args.output_dir and args.output_dir.exists():p.error(f'Outputmappen findes allerede: {args.output_dir}')
    supplied=[args.source,args.target,args.source_answer,args.target_answer]
    if args.case and any(x is not None for x in supplied):p.error('--case kan ikke kombineres med de fire tekstargumenter.')
    if args.case:
        try:args.source,args.target,args.source_answer,args.target_answer=patch.read_case(args.case)
        except (OSError,UnicodeError,ValueError,json.JSONDecodeError) as exc:p.error(f'Casefilen kunne ikke læses: {exc}')
    else:
        defaults=['The capital of France is','The capital of Italy is',' Paris',' Rome']
        args.source,args.target,args.source_answer,args.target_answer=[default if value is None else value for value,default in zip(supplied,defaults)]
    if any(len(x)>300 or any(0xD800<=ord(c)<=0xDFFF for c in x) for x in (args.source,args.target,args.source_answer,args.target_answer)):p.error('Tekster må højst være 300 gyldige tegn.')
    torch.set_num_threads(1)
    try:
        tok=AutoTokenizer.from_pretrained(patch.MODEL,revision=patch.REVISION,local_files_only=args.offline,trust_remote_code=False)
        prepare(tok,args.source,args.target,args.source_answer,args.target_answer)
        model=AutoModelForCausalLM.from_pretrained(patch.MODEL,revision=patch.REVISION,local_files_only=args.offline,trust_remote_code=False,use_safetensors=True).eval()
        result=measure(model,tok,args.source,args.target,args.source_answer,args.target_answer)
    except (OSError,ValueError,RuntimeError) as exc:raise SystemExit(f'Forsøget kunne ikke gennemføres: {exc}') from exc
    result['program_sha256']=hashlib.sha256(Path(__file__).read_bytes()).hexdigest();result['helper_sha256']=hashlib.sha256(Path(patch.__file__).read_bytes()).hexdigest()
    output=args.output_dir or Path('position-map-runs')/datetime.now(timezone.utc).strftime('%Y%m%dT%H%M%S_%fZ')
    try:output.mkdir(parents=True,exist_ok=False)
    except FileExistsError:p.error(f'Outputmappen findes allerede: {output}')
    (output/'position_map.json').write_text(json.dumps(result,ensure_ascii=False,indent=2,allow_nan=False),encoding='utf-8')
    with (output/'positions.csv').open('w',encoding='utf-8',newline='') as f:
        writer=csv.DictWriter(f,fieldnames=['layer','position','patched_difference','effect','identity_logit_max']);writer.writeheader();writer.writerows(result['rows'])
    print(f"{len(result['rows'])} indgreb. Gemt i {output}: position_map.json og positions.csv.")

if __name__=='__main__':main()
