#!/usr/bin/env python3
"""Export two next-token distributions from a fixed Pythia-70M checkpoint."""
import argparse
import hashlib
import json
from pathlib import Path

import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = 'EleutherAI/pythia-70m'
REVISION = 'a39f36b100fe8a5377810d56c3f4789b9c53ac42'
PROMPTS = ('The capital of France is', 'The capital of Italy is')
TARGETS = (' Paris', ' Rome')


def generate(model, tokenizer):
    examples = []
    for prompt, target in zip(PROMPTS, TARGETS):
        ids = tokenizer.encode(prompt, add_special_tokens=False)
        whole = tokenizer.encode(prompt + target, add_special_tokens=False)
        if len(whole) != len(ids) + 1 or whole[:-1] != ids:
            raise ValueError('Svaret er ikke ét token efter prompten.')
        target_id = whole[-1]
        with torch.no_grad():
            logits = model(torch.tensor([ids]), use_cache=False).logits[0, -1]
            probabilities = torch.softmax(logits, dim=-1)
            values, indices = torch.topk(probabilities, 5)
            target_probability = float(probabilities[target_id])
            target_rank = int((logits > logits[target_id]).sum()) + 1
        tokens = [{'id': int(i), 'text': tokenizer.decode([int(i)], clean_up_tokenization_spaces=False), 'probability': float(p)} for p, i in zip(values, indices)]
        examples.append({'prompt': prompt, 'input_ids': ids, 'tokens': tokens,
                         'target': {'id': target_id, 'text': target, 'probability': target_probability, 'rank': target_rank},
                         'other_probability': max(0.0, 1.0 - sum(item['probability'] for item in tokens))})
    return {'format': 'learning-lab-pythia-next-v1', 'model': MODEL, 'revision': REVISION,
            'torch_version': str(torch.__version__), 'transformers_version': transformers.__version__,
            'script_sha256': hashlib.sha256(Path(__file__).read_bytes()).hexdigest(), 'examples': examples}


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument('--output', type=Path, default=Path('next-token-example.json'))
    parser.add_argument('--offline', action='store_true')
    args = parser.parse_args()
    if args.output.exists():
        parser.error('Outputfilen findes allerede; vaelg et andet navn.')
    torch.set_num_threads(1)
    tokenizer = AutoTokenizer.from_pretrained(MODEL, revision=REVISION, local_files_only=args.offline)
    model = AutoModelForCausalLM.from_pretrained(MODEL, revision=REVISION, local_files_only=args.offline).eval()
    result = generate(model, tokenizer)
    with args.output.open('x', encoding='utf-8') as file:
        json.dump(result, file, ensure_ascii=False, indent=2, allow_nan=False)
        file.write('\n')
    print('Gemt:', args.output)


if __name__ == '__main__':
    main()
