translate.py

import argparse
import glob
import math
import os
from typing import Optional

import torch
from accelerate import Accelerator, DistributedType, find_executable_batch_size
from torch.utils.data import DataLoader
from tqdm import tqdm
from transformers import (
    DataCollatorWithPadding,
    PreTrainedTokenizerBase,
)

from dataset import DatasetReader, count_lines
from model import load_model_for_inference


def encode_string(text):
    return text.replace("\r", r"\r").replace("\n", r"\n").replace("\t", r"\t")


def get_dataloader(
    accelerator: Accelerator,
    filename: str,
    tokenizer: PreTrainedTokenizerBase,
    batch_size: int,
    max_length: int,
    prompt: str,
) -> DataLoader:
    dataset = DatasetReader(
        filename=filename,
        tokenizer=tokenizer,
        max_length=max_length,
        prompt=prompt,
    )
    if accelerator.distributed_type == DistributedType.XLA:
        data_collator = DataCollatorWithPadding(
            tokenizer,
            padding="max_length",
            max_length=max_length,
            # label_pad_token_id=tokenizer.pad_token_id,
            return_tensors="pt",
        )
    else:
        data_collator = DataCollatorWithPadding(
            tokenizer,
            padding=True,
            # label_pad_token_id=tokenizer.pad_token_id,
            # max_length=max_length, No need to set max_length here, we already truncate in the preprocess function
            # pad_to_multiple_of=8,
            return_tensors="pt",
        )

    return DataLoader(
        dataset,
        batch_size=batch_size,
        collate_fn=data_collator,
        num_workers=0,  # Disable multiprocessing
    )


def main(
    sentences_path: Optional[str],
    sentences_dir: Optional[str],
    files_extension: str,
    output_path: str,
    source_lang: Optional[str],
    target_lang: Optional[str],
    starting_batch_size: int,
    model_name: str = "facebook/m2m100_1.2B",
    lora_weights_name_or_path: str = None,
    force_auto_device_map: bool = False,
    precision: str = None,
    max_length: int = 256,
    num_beams: int = 4,
    num_return_sequences: int = 1,
    do_sample: bool = False,
    temperature: float = 1.0,
    top_k: int = 50,
    top_p: float = 1.0,
    keep_special_tokens: bool = False,
    keep_tokenization_spaces: bool = False,
    repetition_penalty: float = None,
    prompt: str = None,
    trust_remote_code: bool = False,
):
    accelerator = Accelerator()

    if force_auto_device_map and starting_batch_size >= 64:
        print(
            f"WARNING: You are using a very large batch size ({starting_batch_size}) and the auto_device_map  flag. "
            f"auto_device_map will offload model parameters to the CPU when they don't fit on the GPU VRAM. "
            f"If you use a very large batch size, it will offload a lot of parameters to the CPU and slow down the "
            f"inference. You should consider using a smaller batch size, i.e '--starting_batch_size 8'"
        )

    if sentences_path is None and sentences_dir is None:
        raise ValueError(
            "You must specify either --sentences_path or --sentences_dir. Use --help for more details."
        )

    if sentences_path is not None and sentences_dir is not None:
        raise ValueError(
            "You must specify either --sentences_path or --sentences_dir, not both. Use --help for more details."
        )

    if precision is None:
        quantization = None
        dtype = None
    elif precision == "8" or precision == "4":
        quantization = int(precision)
        dtype = None
    elif precision == "fp16":
        quantization = None
        dtype = "float16"
    elif precision == "bf16":
        quantization = None
        dtype = "bfloat16"
    elif precision == "32":
        quantization = None
        dtype = "float32"
    else:
        raise ValueError(
            f"Precision {precision} not supported. Please choose between 8, 4, fp16, bf16, 32 or None."
        )

    model, tokenizer = load_model_for_inference(
        weights_path=model_name,
        quantization=quantization,
        lora_weights_name_or_path=lora_weights_name_or_path,
        torch_dtype=dtype,
        force_auto_device_map=force_auto_device_map,
        trust_remote_code=trust_remote_code,
    )

    is_translation_model = hasattr(tokenizer, "lang_code_to_id")
    lang_code_to_idx = None

    if (
        is_translation_model
        and (source_lang is None or target_lang is None)
        and "small100" not in model_name
    ):
        raise ValueError(
            f"The model you are using requires a source and target language. "
            f"Please specify them with --source-lang and --target-lang. "
            f"The supported languages are: {tokenizer.lang_code_to_id.keys()}"
        )
    if not is_translation_model and (
        source_lang is not None or target_lang is not None
    ):
        if prompt is None:
            print(
                "WARNING: You are using a model that does not support source and target languages parameters "
                "but you specified them. You probably want to use m2m100/nllb200 for translation or "
                "set --prompt to define the task for you model. "
            )
        else:
            print(
                "WARNING: You are using a model that does not support source and target languages parameters "
                "but you specified them."
            )

    if prompt is not None and "%%SENTENCE%%" not in prompt:
        raise ValueError(
            f"The prompt must contain the %%SENTENCE%% token to indicate where the sentence should be inserted. "
            f"Your prompt: {prompt}"
        )

    if is_translation_model:
        try:
            _ = tokenizer.lang_code_to_id[source_lang]
        except KeyError:
            raise KeyError(
                f"Language {source_lang} not found in tokenizer. Available languages: {tokenizer.lang_code_to_id.keys()}"
            )
        tokenizer.src_lang = source_lang

        try:
            lang_code_to_idx = tokenizer.lang_code_to_id[target_lang]
        except KeyError:
            raise KeyError(
                f"Language {target_lang} not found in tokenizer. Available languages: {tokenizer.lang_code_to_id.keys()}"
            )
        if "small100" in model_name:
            tokenizer.tgt_lang = target_lang
            # We don't need to force the BOS token, so we set is_translation_model to False
            is_translation_model = False

    if model.config.model_type == "seamless_m4t":
        # Loading a seamless_m4t model, we need to set a few things to ensure compatibility

        supported_langs = tokenizer.additional_special_tokens
        supported_langs = [lang.replace("__", "") for lang in supported_langs]

        if source_lang is None or target_lang is None:
            raise ValueError(
                f"The model you are using requires a source and target language. "
                f"Please specify them with --source-lang and --target-lang. "
                f"The supported languages are: {supported_langs}"
            )

        if source_lang not in supported_langs:
            raise ValueError(
                f"Language {source_lang} not found in tokenizer. Available languages: {supported_langs}"
            )
        if target_lang not in supported_langs:
            raise ValueError(
                f"Language {target_lang} not found in tokenizer. Available languages: {supported_langs}"
            )

        tokenizer.src_lang = source_lang

    gen_kwargs = {
        "max_new_tokens": max_length,
        "num_beams": num_beams,
        "num_return_sequences": num_return_sequences,
        "do_sample": do_sample,
        "temperature": temperature,
        "top_k": top_k,
        "top_p": top_p,
    }

    if repetition_penalty is not None:
        gen_kwargs["repetition_penalty"] = repetition_penalty

    if is_translation_model:
        gen_kwargs["forced_bos_token_id"] = lang_code_to_idx

    if model.config.model_type == "seamless_m4t":
        gen_kwargs["tgt_lang"] = target_lang

    if accelerator.is_main_process:
        print(
            f"** Translation **\n"
            f"Input file: {sentences_path}\n"
            f"Sentences dir: {sentences_dir}\n"
            f"Output file: {output_path}\n"
            f"Source language: {source_lang}\n"
            f"Target language: {target_lang}\n"
            f"Force target lang as BOS token: {is_translation_model}\n"
            f"Prompt: {prompt}\n"
            f"Starting batch size: {starting_batch_size}\n"
            f"Device: {str(accelerator.device).split(':')[0]}\n"
            f"Num. Devices: {accelerator.num_processes}\n"
            f"Distributed_type: {accelerator.distributed_type}\n"
            f"Max length: {max_length}\n"
            f"Quantization: {quantization}\n"
            f"Precision: {dtype}\n"
            f"Model: {model_name}\n"
            f"LoRA weights: {lora_weights_name_or_path}\n"
            f"Force auto device map: {force_auto_device_map}\n"
            f"Keep special tokens: {keep_special_tokens}\n"
            f"Keep tokenization spaces: {keep_tokenization_spaces}\n"
        )
        print("** Generation parameters **")
        print("\n".join(f"{k}: {v}" for k, v in gen_kwargs.items()))
        print("\n")

    @find_executable_batch_size(starting_batch_size=starting_batch_size)
    def inference(batch_size, sentences_path, output_path):
        nonlocal \
            model, \
            tokenizer, \
            max_length, \
            gen_kwargs, \
            precision, \
            prompt, \
            is_translation_model

        print(f"Translating {sentences_path} with batch size {batch_size}")

        total_lines: int = count_lines(sentences_path)

        data_loader = get_dataloader(
            accelerator=accelerator,
            filename=sentences_path,
            tokenizer=tokenizer,
            batch_size=batch_size,
            max_length=max_length,
            prompt=prompt,
        )

        model, data_loader = accelerator.prepare(model, data_loader)

        samples_seen: int = 0

        with tqdm(
            total=total_lines,
            desc="Dataset translation",
            leave=True,
            ascii=True,
            disable=(not accelerator.is_main_process),
        ) as pbar, open(output_path, "w", encoding="utf-8") as output_file:
            with torch.no_grad():
                for step, batch in enumerate(data_loader):
                    batch["input_ids"] = batch["input_ids"]
                    batch["attention_mask"] = batch["attention_mask"]

                    generated_tokens = accelerator.unwrap_model(model).generate(
                        **batch,
                        **gen_kwargs,
                    )

                    generated_tokens = accelerator.pad_across_processes(
                        generated_tokens, dim=1, pad_index=tokenizer.pad_token_id
                    )

                    generated_tokens = (
                        accelerator.gather(generated_tokens).cpu().numpy()
                    )

                    tgt_text = tokenizer.batch_decode(
                        generated_tokens,
                        skip_special_tokens=not keep_special_tokens,
                        clean_up_tokenization_spaces=not keep_tokenization_spaces,
                    )
                    if accelerator.is_main_process:
                        if (
                            step
                            == math.ceil(
                                math.ceil(total_lines / batch_size)
                                / accelerator.num_processes
                            )
                            - 1
                        ):
                            tgt_text = tgt_text[
                                : (total_lines * num_return_sequences) - samples_seen
                            ]
                        else:
                            samples_seen += len(tgt_text)

                        print(
                            "\n".join(
                                [encode_string(sentence) for sentence in tgt_text]
                            ),
                            file=output_file,
                        )

                    pbar.update(len(tgt_text) // gen_kwargs["num_return_sequences"])

        print(f"Translation done. Output written to {output_path}\n")

    if sentences_path is not None:
        os.makedirs(os.path.abspath(os.path.dirname(output_path)), exist_ok=True)
        inference(sentences_path=sentences_path, output_path=output_path)

    if sentences_dir is not None:
        print(
            f"Translating all files in {sentences_dir}, with extension {files_extension}"
        )
        os.makedirs(os.path.abspath(output_path), exist_ok=True)
        for filename in glob.glob(
            os.path.join(
                sentences_dir, f"*.{files_extension}" if files_extension else "*"
            )
        ):
            output_filename = os.path.join(output_path, os.path.basename(filename))
            inference(sentences_path=filename, output_path=output_filename)

    print("Translation done.\n")


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Run the translation experiments")
    input_group = parser.add_mutually_exclusive_group(required=True)
    input_group.add_argument(
        "--sentences_path",
        default=None,
        type=str,
        help="Path to a txt file containing the sentences to translate. One sentence per line.",
    )

    input_group.add_argument(
        "--sentences_dir",
        type=str,
        default=None,
        help="Path to a directory containing the sentences to translate. "
        "Sentences must be in  .txt files containing containing one sentence per line.",
    )

    parser.add_argument(
        "--files_extension",
        type=str,
        default="txt",
        help="If sentences_dir is specified, extension of the files to translate. Defaults to txt. "
        "If set to an empty string, we will translate all files in the directory.",
    )

    parser.add_argument(
        "--output_path",
        type=str,
        required=True,
        help="Path to a txt file where the translated sentences will be written. If the input is a directory, "
        "the output will be a directory with the same structure.",
    )

    parser.add_argument(
        "--source_lang",
        type=str,
        default=None,
        required=False,
        help="Source language id. See: supported_languages.md. Required for m2m100 and nllb200",
    )

    parser.add_argument(
        "--target_lang",
        type=str,
        default=None,
        required=False,
        help="Source language id. See: supported_languages.md. Required for m2m100 and nllb200",
    )

    parser.add_argument(
        "--starting_batch_size",
        type=int,
        default=128,
        help="Starting batch size, we will automatically reduce it if we find an OOM error."
        "If you use multiple devices, we will divide this number by the number of devices.",
    )

    parser.add_argument(
        "--model_name",
        type=str,
        default="facebook/m2m100_1.2B",
        help="Path to the model to use. See: https://huggingface.co/models",
    )

    parser.add_argument(
        "--lora_weights_name_or_path",
        type=str,
        default=None,
        help="If the model uses LoRA weights, path to those weights. See: https://github.com/huggingface/peft",
    )

    parser.add_argument(
        "--force_auto_device_map",
        action="store_true",
        help=" Whether to force the use of the auto device map. If set to True, "
        "the model will be split across GPUs and CPU to fit the model in memory. "
        "If set to False, a full copy of the model will be loaded  into each GPU. Defaults to False.",
    )

    parser.add_argument(
        "--max_length",
        type=int,
        default=256,
        help="Maximum number of tokens in the source sentence and generated sentence. "
        "Increase this value to translate longer sentences, at the cost of increasing memory usage.",
    )

    parser.add_argument(
        "--num_beams",
        type=int,
        default=5,
        help="Number of beams for beam search, m2m10 author recommends 5, but it might use too much memory",
    )

    parser.add_argument(
        "--num_return_sequences",
        type=int,
        default=1,
        help="Number of possible translation to return for each sentence (num_return_sequences<=num_beams).",
    )

    parser.add_argument(
        "--precision",
        type=str,
        default=None,
        choices=["bf16", "fp16", "32", "4", "8"],
        help="Precision of the model. bf16, fp16 or 32, 8 , 4 "
        "(4bits/8bits quantification, requires bitsandbytes library: https://github.com/TimDettmers/bitsandbytes). "
        "If None, we will use the torch.dtype of the model weights.",
    )

    parser.add_argument(
        "--do_sample",
        action="store_true",
        help="Use sampling instead of beam search.",
    )

    parser.add_argument(
        "--temperature",
        type=float,
        default=0.8,
        help="Temperature for sampling, value used only if do_sample is True.",
    )

    parser.add_argument(
        "--top_k",
        type=int,
        default=100,
        help="If do_sample is True, will sample from the top k most likely tokens.",
    )

    parser.add_argument(
        "--top_p",
        type=float,
        default=0.75,
        help="If do_sample is True, will sample from the top k most likely tokens.",
    )

    parser.add_argument(
        "--keep_special_tokens",
        action="store_true",
        help="Keep special tokens in the decoded text.",
    )

    parser.add_argument(
        "--keep_tokenization_spaces",
        action="store_true",
        help="Do not clean spaces in the decoded text.",
    )

    parser.add_argument(
        "--repetition_penalty",
        type=float,
        default=None,
        help="Repetition penalty.",
    )

    parser.add_argument(
        "--prompt",
        type=str,
        default=None,
        help="Prompt to use for generation. "
        "It must include the special token %%SENTENCE%% which will be replaced by the sentence to translate.",
    )

    parser.add_argument(
        "--trust_remote_code",
        action="store_true",
        help="If set we will trust remote code in HuggingFace models. This is required for some models.",
    )

    args = parser.parse_args()

    main(
        sentences_path=args.sentences_path,
        sentences_dir=args.sentences_dir,
        files_extension=args.files_extension,
        output_path=args.output_path,
        source_lang=args.source_lang,
        target_lang=args.target_lang,
        starting_batch_size=args.starting_batch_size,
        model_name=args.model_name,
        max_length=args.max_length,
        num_beams=args.num_beams,
        num_return_sequences=args.num_return_sequences,
        precision=args.precision,
        do_sample=args.do_sample,
        temperature=args.temperature,
        top_k=args.top_k,
        top_p=args.top_p,
        keep_special_tokens=args.keep_special_tokens,
        keep_tokenization_spaces=args.keep_tokenization_spaces,
        repetition_penalty=args.repetition_penalty,
        prompt=args.prompt,
        trust_remote_code=args.trust_remote_code,
    )