run_parler_tts_training.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

""" Train Parler-TTS using 🤗 Accelerate"""

import logging
import os
import re
import sys
import time
from multiprocess import set_start_method
from datetime import timedelta

from tqdm import tqdm
from pathlib import Path

import torch
from torch.utils.data import DataLoader

import datasets
from datasets import DatasetDict, Dataset, IterableDataset, concatenate_datasets

from huggingface_hub import HfApi

import transformers
from transformers import AutoFeatureExtractor, AutoTokenizer, HfArgumentParser
from transformers.trainer_pt_utils import LengthGroupedSampler
from transformers.optimization import get_scheduler
from transformers.utils import send_example_telemetry


from accelerate import Accelerator, skip_first_batches
from accelerate.utils import set_seed, AutocastKwargs, InitProcessGroupKwargs, TorchDynamoPlugin
from accelerate.utils.memory import release_memory

from parler_tts import (
    ParlerTTSConfig,
    ParlerTTSForConditionalGeneration,
    build_delay_pattern_mask,
)

from training.utils import (
    get_last_checkpoint,
    rotate_checkpoints,
    log_pred,
    log_metric,
    load_all_codec_checkpoints,
    save_codec_checkpoint,
    get_last_codec_checkpoint_step,
)
from training.arguments import ModelArguments, DataTrainingArguments, ParlerTTSTrainingArguments
from training.data import load_multiple_datasets, DataCollatorParlerTTSWithPadding, DataCollatorEncodecWithPadding
from training.eval import clap_similarity, wer, si_sdr

logger = logging.getLogger(__name__)


def main():
    # See all possible arguments in src/transformers/training_args.py
    # or by passing the --help flag to this script.
    # We now keep distinct sets of args, for a cleaner separation of concerns.

    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, ParlerTTSTrainingArguments))
    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
        # If we pass only one argument to the script and it's the path to a json file,
        # let's parse it to get our arguments.
        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
    else:
        model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
    # information sent is the one passed as arguments along with your Python/PyTorch versions.
    send_example_telemetry("run_parler_tts", model_args, data_args)

    if training_args.dtype == "float16":
        mixed_precision = "fp16"
        torch_dtype = torch.float16
    elif training_args.dtype == "bfloat16":
        mixed_precision = "bf16"
        torch_dtype = torch.bfloat16
    else:
        mixed_precision = "no"
        torch_dtype = torch.float32

    if data_args.pad_to_max_length and (
        data_args.max_duration_in_seconds is None
        or data_args.max_prompt_token_length is None
        or data_args.max_description_token_length is None
    ):
        raise ValueError(
            "`pad_to_max_length` is `True` but one of the following parameters has not been set: `max_duration_in_seconds`, `max_prompt_token_length`, `max_description_token_length`"
        )

    padding = "max_length" if data_args.pad_to_max_length else "longest"

    ####### A. Preparation
    kwargs_handlers = [InitProcessGroupKwargs(timeout=timedelta(minutes=120))]

    accelerator = Accelerator(
        gradient_accumulation_steps=training_args.gradient_accumulation_steps,
        mixed_precision=mixed_precision,
        log_with=training_args.report_to,
        project_dir=training_args.output_dir,
        kwargs_handlers=kwargs_handlers,
    )

    accelerator.init_trackers(
        project_name=data_args.wandb_project,
        config={
            "learning_rate": training_args.learning_rate,
            "model_name_or_path": model_args.model_name_or_path,
            "num_train_epochs": training_args.num_train_epochs,
            "gradient_accumulation_steps": training_args.gradient_accumulation_steps,
            "per_device_train_batch_size": training_args.per_device_train_batch_size,
            "global_batch_size": training_args.per_device_train_batch_size * accelerator.num_processes,
            "mixed_precision": mixed_precision,
            "lr_scheduler_type": training_args.lr_scheduler_type,
            "warmup_steps": training_args.warmup_steps,
            "freeze_text_encoder": model_args.freeze_text_encoder,
            "max_duration_in_seconds": data_args.max_duration_in_seconds,
            "weight_decay": training_args.weight_decay,
            "adam_beta1": training_args.adam_beta1,
            "adam_beta2": training_args.adam_beta2,
            "temperature": model_args.temperature,
        },
        init_kwargs={"wandb": {"name": data_args.wandb_run_name}} if data_args.wandb_run_name else {},
    )

    # Detecting last checkpoint and eventually continue from last checkpoint
    last_checkpoint = None
    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
        last_checkpoint = get_last_checkpoint(training_args.output_dir)
        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
            raise ValueError(
                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                "Use --overwrite_output_dir to overcome."
            )
        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
            logger.info(
                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
            )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        handlers=[logging.StreamHandler(sys.stdout)],
    )
    logger.setLevel(logging.INFO if accelerator.is_main_process else logging.WARN)

    # Log a small summary on each proces
    logger.warning(
        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
        f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
    )

    # Set the verbosity to info of the Transformers logger (on main process only)
    if accelerator.is_local_main_process:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()

    logger.info("Training/evaluation parameters %s", training_args)

    # Set seed before initializing model.
    set_seed(training_args.seed)
    num_workers = data_args.preprocessing_num_workers

    # 1. First, lett's instantiate the feature extractor, tokenizers and model
    # Note for distributed training, the .from_pretrained methods guarantee that only
    # one local process can concurrently download model & vocab.

    # load feature extractor
    feature_extractor = AutoFeatureExtractor.from_pretrained(
        model_args.feature_extractor_name or model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        token=data_args.token,
        trust_remote_code=data_args.trust_remote_code,
    )
    sampling_rate = feature_extractor.sampling_rate

    # load prompt tokenizer
    prompt_tokenizer = AutoTokenizer.from_pretrained(
        model_args.prompt_tokenizer_name or model_args.description_tokenizer_name or model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        token=data_args.token,
        trust_remote_code=data_args.trust_remote_code,
        use_fast=model_args.use_fast_tokenizer,
        padding_side=model_args.prompt_padding_side,
    )

    # load description tokenizer
    description_tokenizer = AutoTokenizer.from_pretrained(
        model_args.description_tokenizer_name or model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        token=data_args.token,
        trust_remote_code=data_args.trust_remote_code,
        use_fast=model_args.use_fast_tokenizer,
    )

    if model_args.use_fast_tokenizer:
        logger.warning(
            "Disabling fast tokenizer warning: https://github.com/huggingface/transformers/blob/main/src/transformers/tokenization_utils_base.py#L3231-L3235"
        )
        prompt_tokenizer.deprecation_warnings["Asking-to-pad-a-fast-tokenizer"] = True
        description_tokenizer.deprecation_warnings["Asking-to-pad-a-fast-tokenizer"] = True

    # 2. Now, let's load the dataset

    if data_args.save_to_disk is not None:
        os.makedirs(data_args.save_to_disk, exist_ok=True)

    # assume that the dataset has been saved to `save_to_disk` if the latter is not empty
    dataset_was_precomputed = len(os.listdir(data_args.save_to_disk)) > 0
    if dataset_was_precomputed:
        with accelerator.local_main_process_first():
            vectorized_datasets = datasets.load_from_disk(data_args.save_to_disk)
    else:
        raw_datasets = DatasetDict()

        columns_to_keep = {
            "target_audio_column_name": data_args.target_audio_column_name,
            "prompt_column_name": data_args.prompt_column_name,
        }
        if data_args.description_column_name is not None:
            columns_to_keep["description_column_name"] = data_args.description_column_name

        if training_args.do_train:
            raw_datasets["train"] = load_multiple_datasets(
                accelerator,
                data_args.train_dataset_name,
                data_args.train_dataset_config_name,
                metadata_dataset_names=data_args.train_metadata_dataset_name,
                splits=data_args.train_split_name,
                dataset_samples=data_args.train_dataset_samples,
                seed=training_args.seed,
                cache_dir=model_args.cache_dir,
                num_proc=data_args.preprocessing_num_workers,
                id_column_name=data_args.id_column_name,
                columns_to_keep=columns_to_keep.values(),
                prompt_column_name=data_args.prompt_column_name,
                audio_column_name=data_args.target_audio_column_name,
                sampling_rate=sampling_rate,
                logger=logger,
                # streaming=data_args.streaming, TODO(SG): optionally enable streaming mode
            )

            for key in columns_to_keep:
                if columns_to_keep[key] not in raw_datasets["train"].column_names:
                    raise ValueError(
                        f"--{key} '{columns_to_keep[key]}' not found in dataset '{data_args.train_dataset_name}'."
                        f" Make sure to set `--{key}` to the correct audio column - one of"
                        f" {', '.join(raw_datasets['train'].column_names)}."
                    )

            if data_args.max_train_samples is not None:
                raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))

        if training_args.do_eval:
            raw_datasets["eval"] = load_multiple_datasets(
                accelerator,
                data_args.eval_dataset_name if data_args.eval_dataset_name else data_args.train_dataset_name,
                data_args.eval_dataset_config_name
                if data_args.eval_dataset_config_name
                else data_args.train_dataset_config_name,
                metadata_dataset_names=data_args.eval_metadata_dataset_name,
                splits=data_args.eval_split_name,
                cache_dir=model_args.cache_dir,
                num_proc=data_args.preprocessing_num_workers,
                id_column_name=data_args.id_column_name,
                columns_to_keep=columns_to_keep.values(),
                prompt_column_name=data_args.prompt_column_name,
                audio_column_name=data_args.target_audio_column_name,
                sampling_rate=sampling_rate,
                logger=logger,
                # streaming=data_args.streaming, TODO(SG): optionally enable streaming mode
            )

            if data_args.max_eval_samples is not None:
                with accelerator.local_main_process_first():
                    raw_datasets["eval"] = (
                        raw_datasets["eval"].shuffle(seed=training_args.seed).select(range(data_args.max_eval_samples))
                    )

    # 3. Next, let's load the config.
    config = ParlerTTSConfig.from_pretrained(
        model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        token=data_args.token,
        trust_remote_code=data_args.trust_remote_code,
    )

    # update pad token id and decoder_start_token_id
    config.decoder.update(
        {
            "cross_attention_implementation_strategy": model_args.cross_attention_implementation_strategy
            if model_args.cross_attention_implementation_strategy is not None
            else None
        }
    )
    config.update(
        {
            "pad_token_id": model_args.pad_token_id if model_args.pad_token_id is not None else config.pad_token_id,
            "decoder_start_token_id": model_args.decoder_start_token_id
            if model_args.decoder_start_token_id is not None
            else config.decoder_start_token_id,
        }
    )

    # create model
    model = ParlerTTSForConditionalGeneration.from_pretrained(
        model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        config=config,
        token=data_args.token,
        trust_remote_code=data_args.trust_remote_code,
        attn_implementation=model_args.attn_implementation,
    )

    # enable gradient checkpointing if necessary
    if training_args.gradient_checkpointing:
        model.gradient_checkpointing_enable()

    # 4. Now we preprocess the datasets including loading the audio, resampling and normalization
    # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
    # so that we just need to set the correct target sampling rate and normalize the input
    # via the `feature_extractor`

    # derive max & min input length for sample rate & max duration
    sampling_rate = feature_extractor.sampling_rate
    max_target_length = int(data_args.max_duration_in_seconds * sampling_rate)
    min_target_length = int(data_args.min_duration_in_seconds * sampling_rate)
    target_audio_column_name = data_args.target_audio_column_name
    description_column_name = data_args.description_column_name
    prompt_column_name = data_args.prompt_column_name
    feature_extractor_input_name = feature_extractor.model_input_names[0]
    audio_encoder_pad_token_id = config.decoder.pad_token_id
    audio_encoder_eos_token_id = config.decoder.eos_token_id
    audio_encoder_bos_token_id = model.generation_config.decoder_start_token_id
    max_length = model.generation_config.max_length
    num_codebooks = model.decoder.config.num_codebooks
    bandwidth = model_args.bandwidth
    attn_implementation = model_args.attn_implementation

    # Freeze Encoders
    model.freeze_encoders(model_args.freeze_text_encoder)

    # Test all gather - used for warmout and avoiding timeout
    logger.debug(str(accelerator.process_index), main_process_only=False, in_order=True)
    test_tensor = torch.tensor([accelerator.process_index], device=accelerator.device)
    gathered_tensor = accelerator.gather(test_tensor)
    print("gathered_tensor", gathered_tensor)
    accelerator.wait_for_everyone()

    if not dataset_was_precomputed:
        # Filter on text length
        if description_column_name is not None and data_args.max_text_length is not None:
            with accelerator.local_main_process_first():
                # filter description that is shorter than max_text_length
                raw_datasets = raw_datasets.filter(
                    lambda x: len(x) < data_args.max_text_length,
                    num_proc=num_workers,
                    input_columns=[description_column_name],
                )

        # Preprocessing the dataset.
        # We need to tokenize the texts.
        def pass_through_processors(description, prompt):
            batch = {}

            batch["input_ids"] = description_tokenizer(description.strip())["input_ids"]
            batch["prompt_input_ids"] = prompt_tokenizer(prompt.strip())["input_ids"]

            return batch

        with accelerator.local_main_process_first():
            # this is a trick to avoid to rewrite the entire audio column which takes ages
            vectorized_datasets = raw_datasets.map(
                pass_through_processors,
                remove_columns=next(iter(raw_datasets.values())).column_names,
                input_columns=[description_column_name, prompt_column_name],
                num_proc=num_workers,
                desc="preprocess datasets",
            )

        # We use Accelerate to perform distributed inference
        # T5 doesn't support fp16
        autocast_kwargs = AutocastKwargs(enabled=(mixed_precision != "fp16"))

        # Now we encode the audio labels with encodec.
        ####### B. Encode audio

        logger.info("*** Encode target audio with encodec ***")

        # no need to prepare audio_decoder because used for inference without mixed precision
        # see: https://huggingface.co/docs/accelerate/main/en/package_reference/accelerator#accelerate.Accelerator.prepare
        if training_args.torch_compile:
            audio_decoder = accelerator.prepare_model(model.audio_encoder, evaluation_mode=True)
        else:
            audio_decoder = model.audio_encoder

        encoder_data_collator = DataCollatorEncodecWithPadding(
            feature_extractor,
            audio_column_name=target_audio_column_name,
            feature_extractor_input_name=feature_extractor_input_name,
            max_length=max_target_length,
            padding=padding,
        )

        def apply_audio_decoder(batch):
            len_audio = batch.pop("len_audio")
            audio_decoder.to(batch["input_values"].device).eval()
            with torch.no_grad():
                labels = audio_decoder.encode(**batch, bandwidth=bandwidth)["audio_codes"]
            output = {}
            output["len_audio"] = len_audio
            # (1, bsz, codebooks, seq_len) -> (bsz, seq_len, codebooks)
            output["labels"] = labels.squeeze(0).transpose(1, 2)

            # if `pad_to_max_length`, the maximum corresponding audio length of the current batch is max_duration*sampling_rate
            max_length = len_audio.max() if padding != "max_length" else max_target_length
            output["ratio"] = torch.ones_like(len_audio) * labels.shape[-1] / max_length
            return output

        # (1, codebooks, seq_len) where seq_len=1
        bos_labels = torch.ones((1, num_codebooks, 1)) * audio_encoder_bos_token_id

        def postprocess_dataset(labels):
            # (1, codebooks, seq_len)
            labels = torch.tensor(labels).unsqueeze(0)
            # add bos
            labels = torch.cat([bos_labels, labels], dim=-1)

            labels, delay_pattern_mask = build_delay_pattern_mask(
                labels,
                bos_token_id=audio_encoder_bos_token_id,
                pad_token_id=audio_encoder_eos_token_id,
                max_length=labels.shape[-1] + num_codebooks,
                num_codebooks=num_codebooks,
            )

            # the first ids of the delay pattern mask are precisely labels, we use the rest of the labels mask
            # to take care of EOS
            # we want labels to look like this:
            #  - [B, a, b, E, E, E, E]
            #  - [B, B, c, d, E, E, E]
            #  - [B, B, B, e, f, E, E]
            #  - [B, B, B, B, g, h, E]
            labels = torch.where(delay_pattern_mask == -1, audio_encoder_eos_token_id, delay_pattern_mask)

            # the first timestamp is associated to a row full of BOS, let's get rid of it
            # we also remove the last timestampts (full of PAD)
            output = {"labels": labels[:, 1:]}
            return output

        for split in vectorized_datasets:
            data_loader = DataLoader(
                raw_datasets[split],
                batch_size=training_args.audio_encoder_per_device_batch_size,
                collate_fn=encoder_data_collator,
                num_workers=training_args.dataloader_num_workers,
                pin_memory=True,
            )
            data_loader = accelerator.prepare(data_loader)
            total_inference_steps = len(data_loader)

            start_step = get_last_codec_checkpoint_step(os.path.join(data_args.temporary_save_to_disk, split))
            accelerator.wait_for_everyone()
            if start_step > 0:
                logger.info(f"Resuming {split} from step {start_step}")
                # efficiently skip the first n batches
                start_step += 1
                data_loader = skip_first_batches(data_loader, start_step)

            all_generated_labels = []
            all_lens = []
            if start_step < total_inference_steps:
                for i, batch in enumerate(tqdm(data_loader, disable=not accelerator.is_local_main_process)):
                    cur_step = start_step + i
                    generate_labels = apply_audio_decoder(batch)
                    generate_labels = accelerator.pad_across_processes(generate_labels, dim=1, pad_index=0)
                    generate_labels = accelerator.gather_for_metrics(generate_labels)

                    if accelerator.is_main_process:
                        lab = generate_labels["labels"].cpu().transpose(1, 2).to(torch.int16)
                        rat = generate_labels["ratio"].cpu().squeeze(1)
                        lens = generate_labels["len_audio"].cpu().squeeze(1)
                        lab = [l[:, : int(ratio * length)] for (l, ratio, length) in zip(lab, rat, lens)]

                        all_generated_labels.extend(lab)
                        all_lens.extend(lens)

                        if ((cur_step + 1) % data_args.save_codec_steps == 0) or (
                            cur_step == total_inference_steps - 1
                        ):
                            tmp_labels = Dataset.from_dict({"labels": all_generated_labels, "target_length": all_lens})
                            tmp_labels = tmp_labels.map(
                                postprocess_dataset,
                                num_proc=data_args.preprocessing_num_workers,  # this one is resource consuming if many processor.
                                input_columns=["labels"],
                                desc="Postprocessing labeling",
                            )
                            save_codec_checkpoint(
                                os.path.join(data_args.temporary_save_to_disk, split), tmp_labels, cur_step
                            )
                            all_generated_labels = []
                            all_lens = []

                accelerator.wait_for_everyone()

            if accelerator.is_main_process and len(all_generated_labels) > 0:
                tmp_labels = Dataset.from_dict({"labels": all_generated_labels, "target_length": all_lens})
                tmp_labels = tmp_labels.map(
                    postprocess_dataset,
                    num_proc=data_args.preprocessing_num_workers,  # this one is resource consuming if many processor.
                    input_columns=["labels"],
                    desc="Postprocessing labeling",
                )
                save_codec_checkpoint(os.path.join(data_args.temporary_save_to_disk, split), tmp_labels, cur_step)
                all_generated_labels = []
                all_lens = []
            accelerator.wait_for_everyone()

            del all_generated_labels
            accelerator.wait_for_everyone()

            with accelerator.local_main_process_first():
                tmp_labels = load_all_codec_checkpoints(os.path.join(data_args.temporary_save_to_disk, split)).select(
                    range(len(vectorized_datasets[split]))
                )
                logger.info(f"Concatenating {split}: {tmp_labels} with {vectorized_datasets[split]}")
                vectorized_datasets[split] = concatenate_datasets([vectorized_datasets[split], tmp_labels], axis=1)

        accelerator.free_memory()
        del generate_labels, all_lens

        with accelerator.local_main_process_first():
            # NOTE: filtering is done at the end because in the `datasets` library, caching audio files is done after most operations
            # caching audio files is time and disk-space consuming, so we want to avoid it at all costs, especially for large (>1Kh) audio datasets.
            # That's also why we avoid to concat the processed datasets (vectorized_datasets) with the audio column present in raw_datasets.

            def is_audio_in_length_range(length):
                return length > min_target_length and length < max_target_length

            # filter data that is shorter than min_target_length
            vectorized_datasets = vectorized_datasets.filter(
                is_audio_in_length_range,
                num_proc=num_workers,
                input_columns=["target_length"],
            )

        if description_column_name is not None and data_args.max_description_token_length is not None:
            with accelerator.local_main_process_first():
                # filter description that is shorter than max_text_length
                vectorized_datasets = vectorized_datasets.filter(
                    lambda x: len(x) < data_args.max_description_token_length,
                    num_proc=num_workers,
                    input_columns=["input_ids"],
                )

        if data_args.max_prompt_token_length is not None:
            with accelerator.local_main_process_first():
                # filter description that is shorter than max_text_length
                vectorized_datasets = vectorized_datasets.filter(
                    lambda x: len(x) < data_args.max_prompt_token_length,
                    num_proc=num_workers,
                    input_columns=["prompt_input_ids"],
                )

    if data_args.save_to_disk is not None and not dataset_was_precomputed:
        if accelerator.is_main_process:
            vectorized_datasets.save_to_disk(
                data_args.save_to_disk,
                num_proc=min(data_args.preprocessing_num_workers, len(vectorized_datasets["eval"]) - 1),
            )
        accelerator.wait_for_everyone()
        logger.info(f"Dataset saved at {data_args.save_to_disk}")

    audio_max_length = None
    if padding == "max_length":
        audio_max_length = max(vectorized_datasets["train"]["target_length"])
        with accelerator.local_main_process_first():
            max_sample = vectorized_datasets["train"].filter(
                lambda x: x == audio_max_length,
                num_proc=num_workers,
                input_columns=["target_length"],
            )
        audio_max_length = max([len(l[0]) for l in max_sample["labels"]])

    if description_column_name is not None and data_args.max_description_token_length is not None:
        with accelerator.local_main_process_first():
            # filter description that is shorter than max_text_length
            vectorized_datasets = vectorized_datasets.filter(
                lambda x: len(x) < data_args.max_description_token_length,
                num_proc=num_workers,
                input_columns=["input_ids"],
            )

    if data_args.max_prompt_token_length is not None:
        with accelerator.local_main_process_first():
            # filter description that is shorter than max_text_length
            vectorized_datasets = vectorized_datasets.filter(
                lambda x: len(x) < data_args.max_prompt_token_length,
                num_proc=num_workers,
                input_columns=["prompt_input_ids"],
            )

    if training_args.group_by_length:
        # apply a simple heuristic to take into account audio and text lengths
        def add_target_lengths(target_length, prompt, description):
            return {"target_length": target_length + len(prompt) + len(description)}

        with accelerator.local_main_process_first():
            vectorized_datasets = vectorized_datasets.map(
                add_target_lengths,
                num_proc=num_workers,
                input_columns=["target_length", "prompt_input_ids", "input_ids"],
            )

    # for large datasets it is advised to run the preprocessing on a
    # single machine first with ``args.preprocessing_only`` since there will mostly likely
    # be a timeout when running the script in distributed mode.
    # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
    # cached dataset
    if data_args.preprocessing_only and data_args.save_to_disk is None:
        raise ValueError(
            "`preprocessing_only=True` but `save_to_disk` is not set. The latter should indicates where to save the dataset locally."
        )
    elif data_args.preprocessing_only:
        logger.info(f"Data preprocessing finished. Files save at {data_args.save_to_disk}")
        return

    # 6. Next, we can prepare the training.

    # Let's use word CLAP similary and WER metrics as our evaluation metrics,
    def compute_metrics(
        audios,
        descriptions,
        prompts,
        device="cpu",
        compute_clap_similarity_metric=False,
        compute_noise_level_metric=False,
        noise_level_to_compute_clean_wer=None,
    ):
        results = {}
        input_ids = descriptions
        texts = description_tokenizer.batch_decode(input_ids, skip_special_tokens=True)
        prompts = prompt_tokenizer.batch_decode(prompts, skip_special_tokens=True)
        audios = [a.float().cpu().numpy() for a in audios]

        if compute_clap_similarity_metric:
            clap_score = clap_similarity(
                model_args.clap_model_name_or_path, texts, audios, device, input_sampling_rate=sampling_rate
            )
            results["clap"] = clap_score

        si_sdr_measures = None
        if compute_noise_level_metric:
            si_sdr_measures = si_sdr(audios, device, input_sampling_rate=sampling_rate)

        word_error, transcriptions, clean_word_error, noisy_word_error, percent_clean_samples = wer(
            model_args.asr_model_name_or_path,
            prompts,
            audios,
            device,
            training_args.per_device_eval_batch_size,
            sampling_rate,
            noise_level_to_compute_clean_wer,
            si_sdr_measures,
        )
        results["wer"] = word_error
        if clean_word_error is not None:
            results["clean_wer"] = clean_word_error
            results["noisy_word_error"] = noisy_word_error
            results["percent_clean_samples"] = percent_clean_samples

        return results, texts, prompts, audios, transcriptions, si_sdr_measures

    # Define Training Schedule
    # Store some constants
    per_device_train_batch_size = int(training_args.per_device_train_batch_size)
    train_batch_size = per_device_train_batch_size * accelerator.num_processes
    gradient_accumulation_steps = int(training_args.gradient_accumulation_steps)
    per_device_eval_batch_size = int(training_args.per_device_eval_batch_size)

    if training_args.max_steps < 0:
        num_epochs = int(training_args.num_train_epochs)
        steps_per_epoch = len(vectorized_datasets["train"]) // (train_batch_size * gradient_accumulation_steps)
        total_train_steps = steps_per_epoch * num_epochs
    elif training_args.max_steps > 0:
        logger.info("max_steps is given, it will override any value given in num_train_epochs")
        total_train_steps = int(training_args.max_steps)
        # Setting a very large number of epochs so we go as many times as necessary over the iterator.
        num_epochs = sys.maxsize
        steps_per_epoch = total_train_steps

    if training_args.eval_steps is None:
        logger.info(f"eval_steps is not set, evaluating at the end of each epoch")
        eval_steps = steps_per_epoch
    else:
        eval_steps = training_args.eval_steps

    # T5 doesn't support fp16
    autocast_kwargs = AutocastKwargs(enabled=(mixed_precision != "fp16"))

    # Define optimizer, LR scheduler, collator
    optimizer = torch.optim.AdamW(
        params=model.parameters(),
        lr=training_args.learning_rate,
        betas=(training_args.adam_beta1, training_args.adam_beta2),
        eps=training_args.adam_epsilon,
        weight_decay=training_args.weight_decay,
    )

    # LR scheduler gets stepped by `num_processes` each time -> account for this in warmup / total steps
    lr_scheduler = get_scheduler(
        name=training_args.lr_scheduler_type,
        optimizer=optimizer,
        num_warmup_steps=training_args.get_warmup_steps(total_train_steps) * accelerator.num_processes,
        num_training_steps=total_train_steps * accelerator.num_processes,
    )

    # Instantiate custom data collator
    data_collator = DataCollatorParlerTTSWithPadding(
        prompt_tokenizer=prompt_tokenizer,
        description_tokenizer=description_tokenizer,
        pad_to_multiple_of=data_args.pad_to_multiple_of,
        padding=padding,
        prompt_max_length=data_args.max_prompt_token_length,
        description_max_length=data_args.max_description_token_length,
        audio_max_length=audio_max_length,
    )

    # Prepare everything with accelerate
    model, optimizer, lr_scheduler = accelerator.prepare(model, optimizer, lr_scheduler)

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {total_train_steps * train_batch_size * gradient_accumulation_steps}")
    logger.info("  Instantaneous batch size per device =" f" {per_device_train_batch_size}")
    logger.info("  Gradient accumulation steps =" f" {gradient_accumulation_steps}")
    logger.info(
        f"  Total train batch size (w. parallel & distributed) = {train_batch_size * gradient_accumulation_steps}"
    )
    logger.info(f"  Total optimization steps = {total_train_steps}")

    # ======================== Training ================================
    train_time = 0
    train_start = time.time()
    steps_trained_progress_bar = tqdm(
        range(total_train_steps), desc="Train steps ... ", position=0, disable=not accelerator.is_local_main_process
    )
    continue_training = True
    epochs_trained = 0
    cur_step = 0

    checkpoint = None
    if training_args.resume_from_checkpoint is not None:
        checkpoint = training_args.resume_from_checkpoint
    elif last_checkpoint is not None:
        checkpoint = last_checkpoint

    if accelerator.is_main_process:
        if training_args.push_to_hub:
            api = HfApi(token=training_args.hub_token)

            # Create repo (repo_name from args or inferred)
            repo_name = training_args.hub_model_id
            if repo_name is None:
                repo_name = Path(training_args.output_dir).absolute().name
            repo_id = api.create_repo(repo_name, exist_ok=True).repo_id

            with open(os.path.join(training_args.output_dir, ".gitignore"), "w+") as gitignore:
                if "wandb" not in gitignore:
                    gitignore.write("wandb\n")
        elif training_args.output_dir is not None:
            os.makedirs(training_args.output_dir, exist_ok=True)
    accelerator.wait_for_everyone()

    # Now save everything to be able to create a single processor later
    # make sure all processes wait until data is saved
    # only the main process saves them
    if accelerator.is_main_process:
        # save feature extractor, tokenizer and config
        if (
            model_args.prompt_tokenizer_name is None
            and model_args.description_tokenizer_name
            or (model_args.prompt_tokenizer_name == model_args.description_tokenizer_name)
        ):
            prompt_tokenizer.save_pretrained(training_args.output_dir)
        else:
            logger.warning(
                f"Prompt tokenizer ('{model_args.prompt_tokenizer_name}') and description tokenizer ('{model_args.description_tokenizer_name}') are not the same. Saving only the prompt tokenizer."
            )
            prompt_tokenizer.save_pretrained(training_args.output_dir)

        feature_extractor.save_pretrained(training_args.output_dir)
        config.save_pretrained(training_args.output_dir)
    accelerator.wait_for_everyone()

    if checkpoint is not None:
        accelerator.load_state(checkpoint)
        # Find num steps and epoch from saved state string pattern
        pattern = r"checkpoint-(\d+)-epoch-(\d+)"
        match = re.search(pattern, checkpoint)
        cur_step = int(match.group(1))
        epochs_trained = int(match.group(2))

        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
        logger.info(f"  Continuing training from epoch {epochs_trained}")
        logger.info(f"  Continuing training from global step {cur_step}")

        steps_trained_progress_bar.update(cur_step)

        for epoch in range(0, epochs_trained):
            with accelerator.local_main_process_first():
                vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed)

        if training_args.max_steps < 0:
            # we know exactly the number of steps per epoch, so can skip through the required number of batches
            resume_step = (cur_step - epochs_trained * steps_per_epoch) * gradient_accumulation_steps
        else:
            # Currently we don't know how many steps we've taken in the current epoch
            # So we just shuffle the dataset one extra time and start from a fresh epoch
            # This is "good enough" for our purposes but not fully correct
            resume_step = None
            with accelerator.local_main_process_first():
                vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed)
    else:
        resume_step = None

    gen_kwargs = {
        "do_sample": model_args.do_sample,
        "temperature": model_args.temperature,
        "max_length": model_args.max_length,
        # Because of the delayed pattern mask, generation might stop earlier because of unexpected behaviour
        # on the first tokens of the codebooks that are delayed.
        # This fix the issue.
        "min_new_tokens": num_codebooks + 1,
    }

    # Define gradient update step fn
    def train_step(
        batch,
        accelerator,
        autocast_kwargs,
    ):
        if mixed_precision == "fp16":
            # fp16 doesn't work with T5-like models
            with accelerator.autocast(autocast_handler=autocast_kwargs):
                if training_args.parallel_mode.value != "distributed":
                    encoder_outputs = model.text_encoder(
                        input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None)
                    )
                else:
                    encoder_outputs = model.module.text_encoder(
                        input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None)
                    )
                # we optionnally project last_hidden_state to avoid recomputing every time
                encoder_hidden_states = encoder_outputs.last_hidden_state
                if (
                    config.text_encoder.hidden_size != config.decoder.hidden_size
                    and config.decoder.cross_attention_hidden_size is None
                ):
                    encoder_hidden_states = (
                        model.enc_to_dec_proj(encoder_hidden_states)
                        if training_args.parallel_mode.value != "distributed"
                        else model.module.enc_to_dec_proj(encoder_hidden_states)
                    )

                if batch.get("attention_mask", None) is not None:
                    encoder_hidden_states = encoder_hidden_states * batch.get("attention_mask", None)[..., None]

                encoder_outputs.last_hidden_state = encoder_hidden_states
                batch["encoder_outputs"] = encoder_outputs

        outputs = model(**batch)
        # CE (data) loss
        ce_loss = outputs.loss

        metrics = {"loss": ce_loss}
        return ce_loss, metrics

    # Define eval fn
    def eval_step(
        batch,
        accelerator,
        autocast_kwargs,
    ):
        eval_model = model if not training_args.torch_compile else model._orig_mod

        if mixed_precision == "fp16":
            # fp16 doesn't work with T5-like models
            with accelerator.autocast(autocast_handler=autocast_kwargs):
                if training_args.parallel_mode.value != "distributed":
                    encoder_outputs = model.text_encoder(
                        input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None)
                    )
                else:
                    encoder_outputs = model.module.text_encoder(
                        input_ids=batch.get("input_ids"), attention_mask=batch.get("attention_mask", None)
                    )
                # we optionnally project last_hidden_state to avoid recomputing every time
                encoder_hidden_states = encoder_outputs.last_hidden_state
                if (
                    config.text_encoder.hidden_size != config.decoder.hidden_size
                    and config.decoder.cross_attention_hidden_size is None
                ):
                    encoder_hidden_states = (
                        model.enc_to_dec_proj(encoder_hidden_states)
                        if training_args.parallel_mode.value != "distributed"
                        else model.module.enc_to_dec_proj(encoder_hidden_states)
                    )

                if batch.get("attention_mask", None) is not None:
                    encoder_hidden_states = encoder_hidden_states * batch.get("attention_mask", None)[..., None]

                encoder_outputs.last_hidden_state = encoder_hidden_states
                batch["encoder_outputs"] = encoder_outputs

        with torch.no_grad():
            outputs = eval_model(**batch)
        # CE (data) loss
        ce_loss = outputs.loss
        metrics = {"loss": ce_loss}
        return metrics

    def generate_step(batch, accelerator):
        batch.pop("decoder_attention_mask", None)
        eval_model = accelerator.unwrap_model(model, keep_fp32_wrapper=True)
        if training_args.torch_compile:
            # if the model is compiled, we use the original model bc compile is not compatible with .generate
            eval_model = model._orig_mod

        # since we've might have loaded the weights in fp32, we have to autocast to ensure FA2 weights are in half-precision.
        # with accelerator.autocast(autocast_handler=AutocastKwargs(enabled=(attn_implementation=="flash_attention_2"))):
        output_audios = eval_model.generate(**batch, **gen_kwargs)
        output_audios = accelerator.pad_across_processes(output_audios, dim=1, pad_index=0)
        return output_audios

    model.train()

    for epoch in range(epochs_trained, num_epochs):
        with accelerator.local_main_process_first():
            vectorized_datasets["train"] = vectorized_datasets["train"].shuffle(training_args.seed)
        sampler = None
        if training_args.group_by_length:
            sampler = LengthGroupedSampler(train_batch_size, lengths=vectorized_datasets["train"]["target_length"])
        train_dataloader = DataLoader(
            vectorized_datasets["train"],
            collate_fn=data_collator,
            batch_size=per_device_train_batch_size,
            sampler=sampler,
            num_workers=training_args.dataloader_num_workers,
            pin_memory=training_args.dataloader_pin_memory,
        )
        train_dataloader = accelerator.prepare(train_dataloader)
        if hasattr(train_dataloader, "dataset") and isinstance(train_dataloader.dataset, IterableDataset):
            train_dataloader.dataset.set_epoch(epoch)

        if resume_step is not None:
            # Skip the first N batches in the dataloader when resuming from a checkpoint
            logger.info(f"  Skip first {resume_step} batches")
            train_dataloader = accelerator.skip_first_batches(train_dataloader, resume_step)
            resume_step = None
            accelerator.wait_for_everyone()

        for batch in train_dataloader:
            with accelerator.accumulate(model):
                loss, train_metric = train_step(batch, accelerator, autocast_kwargs)
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    accelerator.clip_grad_norm_(model.parameters(), training_args.max_grad_norm)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()

            # Check if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                steps_trained_progress_bar.update(1)
                cur_step += 1

                if cur_step % training_args.logging_steps == 0:
                    steps_trained_progress_bar.write(
                        f"Step... ({cur_step} / {total_train_steps} | Loss:"
                        f" {train_metric['loss']}, Learning Rate:"
                        f" {lr_scheduler.get_last_lr()[0]})"
                    )
                    log_metric(
                        accelerator,
                        metrics=train_metric,
                        learning_rate=lr_scheduler.get_last_lr()[0],
                        train_time=train_time + time.time() - train_start,
                        step=cur_step,
                        epoch=epoch,
                        prefix="train",
                    )

                # save checkpoint and weights after each save_steps and at the end of training
                if (cur_step % training_args.save_steps == 0) or cur_step == total_train_steps:
                    intermediate_dir = os.path.join(training_args.output_dir, f"checkpoint-{cur_step}-epoch-{epoch}")
                    # safe_serialization=False to avoid shared tensors saving issue (TODO(YL): it's a temporary fix)
                    # https://github.com/huggingface/transformers/issues/27293#issuecomment-1872560074
                    accelerator.save_state(output_dir=intermediate_dir, safe_serialization=False)
                    accelerator.wait_for_everyone()
                    if accelerator.is_main_process:
                        rotate_checkpoints(
                            training_args.save_total_limit, output_dir=training_args.output_dir, logger=logger
                        )

                        if cur_step == total_train_steps:
                            # un-wrap student model for save
                            unwrapped_model = accelerator.unwrap_model(model)
                            unwrapped_model.save_pretrained(training_args.output_dir)

                        if training_args.push_to_hub:
                            api.upload_folder(
                                repo_id=repo_id,
                                folder_path=training_args.output_dir,
                                commit_message=f"Saving train state of step {cur_step}",
                                run_as_future=True,
                            )
                    accelerator.wait_for_everyone()

                if training_args.do_eval and (cur_step % eval_steps == 0 or cur_step == total_train_steps):
                    train_time += time.time() - train_start
                    # ======================== Evaluating ==============================
                    model.eval()
                    eval_metrics = []
                    eval_preds = []
                    eval_descriptions = []
                    eval_prompts = []
                    eval_start = time.time()

                    # release training input batch
                    batch = release_memory(batch)

                    validation_dataloader = DataLoader(
                        vectorized_datasets["eval"],
                        collate_fn=data_collator,
                        batch_size=per_device_eval_batch_size,
                        drop_last=False,
                        num_workers=training_args.eval_dataloader_num_workers,
                        pin_memory=training_args.dataloader_pin_memory,
                    )
                    validation_dataloader = accelerator.prepare(validation_dataloader)

                    for batch in tqdm(
                        validation_dataloader,
                        desc=f"Evaluating - Inference ...",
                        position=2,
                        disable=not accelerator.is_local_main_process,
                    ):
                        # Model forward
                        eval_metric = eval_step(batch, accelerator, autocast_kwargs)
                        eval_metric = accelerator.gather_for_metrics(eval_metric)
                        eval_metric = {key: val.unsqueeze(0) if val.ndim == 0 else val for (key,val) in eval_metric.items()}
                        eval_metrics.append(eval_metric)

                    if training_args.predict_with_generate:
                        validation_dataloader = DataLoader(
                            vectorized_datasets["eval"],
                            collate_fn=data_collator,
                            batch_size=per_device_eval_batch_size,
                            drop_last=False,
                            num_workers=training_args.dataloader_pin_memory,
                            pin_memory=training_args.dataloader_pin_memory,
                        )
                        validation_dataloader = accelerator.prepare(validation_dataloader)
                        # generation
                        for batch in tqdm(
                            validation_dataloader,
                            desc=f"Evaluating - Generation ...",
                            position=2,
                            disable=not accelerator.is_local_main_process,
                        ):
                            generated_audios = generate_step(batch, accelerator)
                            # Gather all predictions and targets
                            generated_audios, input_ids, prompts = accelerator.pad_across_processes(
                                (generated_audios, batch["input_ids"], batch["prompt_input_ids"]), dim=1, pad_index=0
                            )
                            generated_audios, input_ids, prompts = accelerator.gather_for_metrics(
                                (generated_audios, input_ids, prompts)
                            )
                            eval_preds.extend(generated_audios.to("cpu"))
                            eval_descriptions.extend(input_ids.to("cpu"))
                            eval_prompts.extend(prompts.to("cpu"))

                    eval_time = time.time() - eval_start
                    # normalize eval metrics
                    eval_metrics = {
                        key: torch.mean(torch.cat([d[key] for d in eval_metrics])).to("cpu") for key in eval_metrics[0]
                    }

                    # compute metrics
                    metrics_desc = ""
                    if training_args.predict_with_generate:
                        if accelerator.is_local_main_process:
                            (
                                metric_values,
                                pred_descriptions,
                                pred_prompts,
                                audios,
                                transcriptions,
                                si_sdr_measures,
                            ) = compute_metrics(
                                eval_preds,
                                eval_descriptions,
                                eval_prompts,
                                accelerator.device,
                                training_args.compute_clap_similarity_metric,
                                training_args.compute_noise_level_metric,
                                training_args.noise_level_to_compute_clean_wer,
                            )
                            eval_metrics.update(metric_values)
                            metrics_desc = " ".join([f"Eval {key}: {value} |" for key, value in metric_values.items()])
                            if "wandb" in training_args.report_to:
                                log_pred(
                                    accelerator,
                                    pred_descriptions,
                                    pred_prompts,
                                    transcriptions,
                                    audios,
                                    si_sdr_measures,
                                    sampling_rate=sampling_rate,
                                    step=cur_step,
                                    prefix="eval",
                                )
                        accelerator.wait_for_everyone()

                    # Print metrics and update progress bar
                    if accelerator.is_local_main_process:
                        steps_trained_progress_bar.write(
                            f"Eval results for step ({cur_step} / {total_train_steps} | Eval Loss: {eval_metrics['loss']} |"
                            f" {metrics_desc})"
                        )

                    log_metric(
                        accelerator,
                        metrics=eval_metrics,
                        train_time=eval_time,
                        step=cur_step,
                        epoch=epoch,
                        prefix="eval",
                    )

                    # release eval batch and relax metrics
                    eval_metrics, eval_preds, eval_descriptions, eval_prompts, batch, eval_metric = release_memory(
                        eval_metrics, eval_preds, eval_descriptions, eval_prompts, batch, eval_metric
                    )
                    if training_args.predict_with_generate:
                        generated_audios, input_ids, prompts = release_memory(generated_audios, input_ids, prompts)

                    # train mode
                    model.train()

                    # flush the train metrics
                    train_start = time.time()

                # break condition
                if cur_step == total_train_steps:
                    continue_training = False
                    break

        if not continue_training:
            break

    accelerator.end_training()


if __name__ == "__main__":
    main()