Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

10.1 KB
Raw Blame
#!/usr/bin/env python3
# -*- encoding: utf-8 -*-

import os
import sys
import torch
import torch.nn as nn
import hydra
import logging
import time
import argparse
from io import BytesIO

from contextlib import nullcontext
import torch.distributed as dist

from omegaconf import DictConfig, OmegaConf
from funasr.utils.amp import autocast, GradScaler
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.algorithms.join import Join
from torch.distributed.fsdp.sharded_grad_scaler import ShardedGradScaler
from funasr.train_utils.average_nbest_models import average_checkpoints

from funasr.register import tables
from funasr.optimizers import optim_classes
from funasr.train_utils.trainer_ds import Trainer
from funasr.schedulers import scheduler_classes
from funasr.train_utils.initialize import initialize
from funasr.download.download_model_from_hub import download_model
from funasr.models.lora.utils import mark_only_lora_as_trainable
from funasr.train_utils.set_all_random_seed import set_all_random_seed
from funasr.train_utils.device_funcs import (
    clear_accelerator_cache,
    get_default_dist_backend,
    set_accelerator_device,
)
from funasr.train_utils.load_pretrained_model import load_pretrained_model
from funasr.utils.misc import prepare_model_dir
from funasr.train_utils.model_summary import model_summary
from funasr import AutoModel

try:
    import deepspeed
except:
    deepspeed = None

_DISTRIBUTED_TRAIN_CONF_KEYS = (
    "use_ddp",
    "use_fsdp",
    "use_deepspeed",
    "deepspeed_config",
)


def _resolve_distributed_config(kwargs, world_size):
    """Resolve distributed settings with top-level values taking precedence."""
    train_conf = dict(kwargs.get("train_conf") or {})

    def get_setting(name, default):
        if name in kwargs:
            return kwargs[name]
        return train_conf.get(name, default)

    use_fsdp = get_setting("use_fsdp", False)
    use_deepspeed = get_setting("use_deepspeed", False)
    deepspeed_config = get_setting("deepspeed_config", "")
    if use_deepspeed and use_fsdp:
        raise ValueError("use_deepspeed and use_fsdp cannot be enabled at the same time")

    trainer_conf = {
        key: value
        for key, value in train_conf.items()
        if key not in _DISTRIBUTED_TRAIN_CONF_KEYS
    }
    use_ddp = world_size > 1 and not use_deepspeed and not use_fsdp
    return use_ddp, use_fsdp, use_deepspeed, deepspeed_config, trainer_conf


@hydra.main(config_name=None, version_base=None)
def main_hydra(kwargs: DictConfig):
    """Main hydra.
    
        Args:
            kwargs: Additional keyword arguments.
        """
    if kwargs.get("debug", False):
        import pdb

        pdb.set_trace()

    assert "model" in kwargs
    if "model_conf" not in kwargs:
        logging.info("download models from model hub: {}".format(kwargs.get("hub", "ms")))
        kwargs = download_model(is_training=kwargs.get("is_training", True), **kwargs)

    main(**kwargs)


def main(**kwargs):

    # set random seed
    """Main.
    
        Args:
            **kwargs: Additional keyword arguments.
        """
    set_all_random_seed(kwargs.get("seed", 0))
    torch.backends.cudnn.enabled = kwargs.get("cudnn_enabled", torch.backends.cudnn.enabled)
    torch.backends.cudnn.benchmark = kwargs.get("cudnn_benchmark", torch.backends.cudnn.benchmark)
    torch.backends.cudnn.deterministic = kwargs.get("cudnn_deterministic", True)
    # open tf32
    torch.backends.cuda.matmul.allow_tf32 = kwargs.get("enable_tf32", True)

    rank = int(os.environ.get("RANK", 0))
    local_rank = int(os.environ.get("LOCAL_RANK", 0))
    world_size = int(os.environ.get("WORLD_SIZE", 1))

    if local_rank == 0:
        tables.print()

    (
        use_ddp,
        use_fsdp,
        use_deepspeed,
        deepspeed_config,
        trainer_conf,
    ) = _resolve_distributed_config(kwargs, world_size)
    dist_backend = kwargs.get("backend", get_default_dist_backend())
    if use_deepspeed:
        logging.info(f"use_deepspeed: {use_deepspeed}")
        deepspeed.init_distributed(dist_backend=dist_backend)
    elif use_ddp or use_fsdp:
        logging.info(f"use_ddp: {use_ddp}, use_fsdp: {use_fsdp}")
        dist.init_process_group(
            backend=dist_backend,
            init_method="env://",
        )
        set_accelerator_device(local_rank)

    # rank = dist.get_rank()

    logging.info("Build model, frontend, tokenizer")
    device = kwargs.get("device", "cuda")
    kwargs["device"] = "cpu"
    model = AutoModel(**kwargs)

    # save config.yaml
    if rank == 0:
        prepare_model_dir(**kwargs)

    # parse kwargs
    kwargs = model.kwargs
    kwargs["device"] = device
    tokenizer = kwargs["tokenizer"]
    frontend = kwargs["frontend"]
    model = model.model
    del kwargs["model"]

    # freeze_param
    freeze_param = kwargs.get("freeze_param", None)
    if freeze_param is not None:
        if "," in freeze_param:
            freeze_param = freeze_param.split(",")
        if not isinstance(freeze_param, (list, tuple)):
            freeze_param = (freeze_param,)
        logging.info("freeze_param is not None: %s", freeze_param)
        for t in freeze_param:
            for k, p in model.named_parameters():
                if k.startswith(t + ".") or k == t:
                    logging.info(f"Setting {k}.requires_grad = False")
                    p.requires_grad = False
    lora_only = kwargs.get("lora_only", False)
    if lora_only:
        lora_bias = kwargs.get("lora_bias", "none")
        logging.info("Enable LoRA-only training with bias=%s", lora_bias)
        mark_only_lora_as_trainable(model, bias=lora_bias)
    if local_rank == 0:
        logging.info(f"{model_summary(model)}")

    trainer = Trainer(
        rank=rank,
        local_rank=local_rank,
        world_size=world_size,
        use_ddp=use_ddp,
        use_fsdp=use_fsdp,
        use_deepspeed=use_deepspeed,
        deepspeed_config=deepspeed_config,
        device=kwargs["device"],
        excludes=kwargs.get("excludes", None),
        output_dir=kwargs.get("output_dir", "./exp"),
        **trainer_conf,
    )

    model = trainer.warp_model(model, **kwargs)

    kwargs["device"] = int(os.environ.get("LOCAL_RANK", 0))
    trainer.device = int(os.environ.get("LOCAL_RANK", 0))

    model, optim, scheduler = trainer.warp_optim_scheduler(model, **kwargs)

    # dataset
    logging.info("Build dataloader")
    dataloader_class = tables.dataloader_classes.get(
        kwargs["dataset_conf"].get("dataloader", "DataloaderMapStyle")
    )
    dataloader = dataloader_class(**kwargs)
    # dataloader_tr, dataloader_val = dataloader_class(**kwargs)

    scaler = GradScaler(enabled=True) if trainer.use_fp16 else None
    scaler = ShardedGradScaler(enabled=trainer.use_fp16) if trainer.use_fsdp else scaler

    trainer.resume_checkpoint(
        model=model,
        optim=optim,
        scheduler=scheduler,
        scaler=scaler,
    )

    early_stopping_patience = kwargs.get("train_conf", {}).get("early_stopping_patience", 0)
    best_val_loss = float("inf")
    epochs_no_improve = 0

    dataloader_tr, dataloader_val = None, None
    for epoch in range(trainer.start_epoch, trainer.max_epoch):
        time1 = time.perf_counter()

        for data_split_i in range(trainer.start_data_split_i, dataloader.data_split_num):
            time_slice_i = time.perf_counter()

            dataloader_tr, dataloader_val = dataloader.build_iter(
                epoch, data_split_i=data_split_i, start_step=trainer.start_step
            )

            trainer.train_epoch(
                model=model,
                optim=optim,
                scheduler=scheduler,
                scaler=scaler,
                dataloader_train=dataloader_tr,
                dataloader_val=dataloader_val,
                epoch=epoch,
                data_split_i=data_split_i,
                data_split_num=dataloader.data_split_num,
                start_step=trainer.start_step,
            )
            trainer.start_step = 0

            device = next(model.parameters()).device
            clear_accelerator_cache(device)

            time_escaped = (time.perf_counter() - time_slice_i) / 3600.0
            logging.info(
                f"\n\nrank: {local_rank}, "
                f"time_escaped_epoch: {time_escaped:.3f} hours, "
                f"estimated to finish {dataloader.data_split_num} data_slices, remaining: {dataloader.data_split_num-data_split_i} slices, {(dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours, "
                f"epoch: {trainer.max_epoch - epoch} epochs, {((trainer.max_epoch - epoch - 1)*dataloader.data_split_num + dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours\n"
            )

        trainer.start_data_split_i = 0
        trainer.validate_epoch(model=model, dataloader_val=dataloader_val, epoch=epoch + 1)
        current_val = trainer.val_loss_avg

        if current_val < best_val_loss:
            logging.info(f"current_val: {current_val}, best_val_loss: {best_val_loss}")
            best_val_loss = current_val
            epochs_no_improve = 0
        else:
            epochs_no_improve += 1
            logging.info(f"No val_loss improvement for {epochs_no_improve}/{early_stopping_patience} epochs")
        if early_stopping_patience > 0 and epochs_no_improve >= early_stopping_patience:
            logging.info(f"Early stopping triggered at epoch {epoch+1}")
            break

        trainer.step_in_epoch = 0
        trainer.save_checkpoint(
            epoch + 1, model=model, optim=optim, scheduler=scheduler, scaler=scaler
        )

        time2 = time.perf_counter()
        time_escaped = (time2 - time1) / 3600.0
        logging.info(
            f"\n\nrank: {local_rank}, "
            f"time_escaped_epoch: {time_escaped:.3f} hours, "
            f"estimated to finish {trainer.max_epoch} "
            f"epoch: {(trainer.max_epoch - epoch) * time_escaped:.3f} hours\n"
        )
        trainer.train_acc_avg = 0.0
        trainer.train_loss_avg = 0.0

    if trainer.rank == 0:
        average_checkpoints(
            trainer.output_dir, trainer.avg_nbest_model, use_deepspeed=trainer.use_deepspeed
        )

    trainer.close()


if __name__ == "__main__":
    main_hydra()