#!/usr/bin/env python3
"""Export a small, reproducible selection of Pythia's input embeddings."""
import argparse
import json
from pathlib import Path

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = 'EleutherAI/pythia-70m'
REVISION = 'a39f36b100fe8a5377810d56c3f4789b9c53ac42'
EXAMPLES = [' cat', ' dog', ' Paris', ' Rome', ' the', 'cat']


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('--offline', action='store_true')
    parser.add_argument('--output', type=Path, required=True)
    parser.add_argument('--tekst', help='Egen tekst til en JSON-fil, højst 200 tegn og 2–16 tokens')
    parser.add_argument('--lag', action='store_true', help='Gem også aktiveringer efter hvert transformerlag')
    args = parser.parse_args()
    if args.lag and args.tekst is None:
        parser.error('--lag kræver --tekst.')
    if args.tekst is not None and (not args.tekst.strip() or len(args.tekst) > 200):
        parser.error('--tekst skal indeholde 1–200 tegn og må ikke være kun mellemrum.')
    torch.set_num_threads(1)
    tokenizer = AutoTokenizer.from_pretrained(MODEL, revision=REVISION, local_files_only=args.offline, trust_remote_code=False)
    if args.tekst is None:
        tokens = []
        for text in EXAMPLES:
            ids = tokenizer(text, add_special_tokens=False).input_ids
            if len(ids) != 1:
                raise ValueError(f'{text!r} is not one token: {ids}')
            tokens.append((text, ids[0]))
    else:
        ids = tokenizer(args.tekst, add_special_tokens=False).input_ids
        if not 2 <= len(ids) <= 16:
            parser.error(f'Teksten gav {len(ids)} tokens; vælg en tekst med 2–16 tokens.')
        tokens = list(zip(tokenizer.convert_ids_to_tokens(ids), ids))
    model = AutoModelForCausalLM.from_pretrained(MODEL, revision=REVISION, local_files_only=args.offline, trust_remote_code=False, use_safetensors=True)
    model.eval()
    with torch.no_grad():
        rows = [{'text': text, 'id': token_id, 'vector': model.get_input_embeddings().weight[token_id].float().tolist()} for text, token_id in tokens]
    data = {'model': MODEL, 'revision': REVISION, 'dimension': len(rows[0]['vector']), 'rows': rows}
    if args.tekst is not None:
        data['format'] = 'learning-lab-pythia-embeddings-v1'
    if args.lag:
        # Capture each block's residual-stream output before the final model layer norm.
        activations, handles = [], []
        for block in model.gpt_neox.layers:
            def capture(_module, _inputs, output):
                activations.append(output[0].detach().float().cpu())
            handles.append(block.register_forward_hook(capture))
        try:
            with torch.no_grad():
                model(input_ids=torch.tensor([ids]), use_cache=False)
        finally:
            for handle in handles:
                handle.remove()
        if len(activations) != 6 or any(tuple(layer.shape) != (1, len(ids), 512) for layer in activations):
            raise RuntimeError('Uventet antal eller form af lagaktiveringer.')
        for position, row in enumerate(rows):
            row['layers'] = [layer[0, position].tolist() for layer in activations]
        data['layer_count'] = len(activations)
    # Refuse accidental overwrite.
    with args.output.open('x', encoding='utf-8') as file:
        if args.tekst is None:
            file.write('// Generated by open-model-lab/export_embeddings.py from pinned model weights.\n')
            file.write('window.PythiaEmbeddingExamples = ' + json.dumps(data, ensure_ascii=False, separators=(',', ':')) + ';\n')
        else:
            json.dump(data, file, ensure_ascii=False, separators=(',', ':'))
    print(f"Saved {len(rows)} token positions of dimension {data['dimension']} to {args.output}")


if __name__ == '__main__':
    main()
