convert_to_ct2.py

import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"

from ctranslate2.converters import TransformersConverter
from transformers import AutoTokenizer

model_dir = "models/translit"
output_dir = "ct2_model"

print("Starting CTranslate2 conversion...")

# Force-load slow SentencePiece tokenizer once to avoid tiktoken path
_ = AutoTokenizer.from_pretrained(model_dir, use_fast=False)

converter = TransformersConverter(model_dir)

converter.convert(
    output_dir,
    quantization="int8",
    force=True,
)

print("✅ Conversion complete! Saved to:", output_dir)