convert_hf_to_ct2.py

import os
from transformers import WhisperTokenizer, WhisperConfig
import argparse
import ctranslate2
import numpy as np

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--model", type=str,  )
    parser.add_argument("--output_dir", type=str,  )
    parser.add_argument("--quantization", default = "int8_float16", help="can be int8, float16 and int8_float16, if it is not set (None), then do not apply quantization"  )
    parser.add_argument("--load_as_float16", action="store_true" )
    parser.add_argument("--low_cpu_mem_usage", action="store_true" )
    parser.add_argument("--trust_remote_code", action="store_true" )
    parser.add_argument("--force", action="store_true" )
    
    args = parser.parse_args()
        
    converter = ctranslate2.converters.TransformersConverter(
        model_name_or_path = args.model,
        load_as_float16 = args.load_as_float16,
        low_cpu_mem_usage = args.low_cpu_mem_usage,
        trust_remote_code = args.trust_remote_code
    )
    
    converter.convert(
        output_dir = args.output_dir, 
        quantization = args.quantization, 
        force = args.force
    )
    
    ## save original tokenizer
    tokenizer = WhisperTokenizer.from_pretrained( args.model )
    tokenizer.save_pretrained( args.output_dir + "/hf_model" )
    
    ## save original model configuration
    config = WhisperConfig.from_pretrained( args.model )
    config.save_pretrained( args.output_dir + "/hf_model" )