scripts/train_f5.py

#!/usr/bin/env python3
"""Cross-platform launcher for F5-TTS fine-tuning.

Auto-detects the compute device (CUDA / MPS / CPU) and adjusts settings:
- bnb_optimizer is only enabled on CUDA (bitsandbytes doesn't support MPS/CPU).
- num_workers is forced to 0 on Windows & macOS (soundfile/LibsndfileError).
- On Linux, num_workers defaults to 12 (or override via F5_NUM_WORKERS env var).
- Windows WMI cache is preloaded to avoid torch import hangs.
"""

from __future__ import annotations

import os
import platform
import socket
import sys

# ---------------------------------------------------------------------------
# Windows-only: preload WMI cache to avoid hang during torch import
# ---------------------------------------------------------------------------
if sys.platform == "win32":
    platform._uname_cache = platform.uname_result(  # type: ignore[attr-defined]
        "Windows", socket.gethostname(), "10", "Windows", "AMD64"
    )

# ---------------------------------------------------------------------------
# Device-safe bnb_optimizer patch
# bitsandbytes only supports CUDA.  On MPS or CPU we silently fall back to
# standard AdamW so the same CLI flags work everywhere.
# ---------------------------------------------------------------------------
import f5_tts.model.trainer as _trainer_module

_original_trainer_init = _trainer_module.Trainer.__init__


def _patched_trainer_init(self, *args, **kwargs):
    bnb = kwargs.get("bnb_optimizer", False)
    if bnb:
        import torch

        if not torch.cuda.is_available():
            kwargs["bnb_optimizer"] = False
            dev = "mps" if torch.backends.mps.is_available() else "cpu"
            print(
                f"\n[Device] CUDA not available (detected {dev}). "
                f"Disabling --bnb_optimizer; using standard AdamW.\n"
            )
    return _original_trainer_init(self, *args, **kwargs)


_trainer_module.Trainer.__init__ = _patched_trainer_init

# ---------------------------------------------------------------------------
# Platform-aware DataLoader patches
# ---------------------------------------------------------------------------
_original_train = _trainer_module.Trainer.train


def _patched_train(self, train_dataset, num_workers=12, resumable_with_seed=None):
    env_workers = os.getenv("F5_NUM_WORKERS")
    if env_workers is not None:
        num_workers = int(env_workers)
        print(f"\n[Workers] Using F5_NUM_WORKERS={num_workers}.\n")
    elif sys.platform == "win32":
        print(
            "\n[Windows] Forcing num_workers=0 to avoid soundfile/LibsndfileError.\n"
        )
        num_workers = 0
    elif sys.platform == "darwin":
        print(
            "\n[macOS] Forcing num_workers=0 for DataLoader stability on MPS.\n"
        )
        num_workers = 0
    else:
        # Linux — keep the default (12) unless overridden
        print(f"\n[Linux] Using num_workers={num_workers}.\n")

    return _original_train(
        self, train_dataset, num_workers=num_workers, resumable_with_seed=resumable_with_seed
    )


_trainer_module.Trainer.train = _patched_train

import torch.utils.data as _data_module

_original_dataloader_init = _data_module.DataLoader.__init__


def _patched_dataloader_init(self, *args, **kwargs):
    # persistent_workers=True is only safe when num_workers > 0.
    # On Windows/macOS we force num_workers=0, so disable it there.
    if kwargs.get("num_workers", 0) == 0:
        kwargs["persistent_workers"] = False
    return _original_dataloader_init(self, *args, **kwargs)


_data_module.DataLoader.__init__ = _patched_dataloader_init

# ---------------------------------------------------------------------------
# Delegate to F5-TTS CLI
# ---------------------------------------------------------------------------
from f5_tts.train.finetune_cli import main

if __name__ == "__main__":
    sys.exit(main())