scripts/export_inference_checkpoint.py

from __future__ import annotations

import argparse
from pathlib import Path

import torch

DEFAULT_SOURCE = Path("/fs/nexus-scratch/psando/nanotts-05-10/gpt2/ckpt_025000.pt")
stem = DEFAULT_SOURCE.stem
DEFAULT_OUTPUT = Path(f"checkpoints/{stem}_inference.pt")

def main() -> None:
    parser = argparse.ArgumentParser(description="Export a slim nanoTTS inference checkpoint.")
    parser.add_argument("--source", type=Path, default=DEFAULT_SOURCE)
    parser.add_argument("--output", type=Path, default=DEFAULT_OUTPUT)
    parser.add_argument("--overwrite", action="store_true")
    args = parser.parse_args()

    if not args.source.exists():
        raise FileNotFoundError(f"Source checkpoint not found: {args.source}")
    if args.output.exists() and not args.overwrite:
        raise FileExistsError(f"Output already exists: {args.output}. Pass --overwrite to replace it.")

    checkpoint = torch.load(args.source, map_location="cpu")
    inference_checkpoint = {
        "model": checkpoint["model"],
        "model_args": checkpoint["model_args"],
        "iter_num": checkpoint.get("iter_num"),
        "train_loss": checkpoint.get("train_loss"),
        "val_loss": checkpoint.get("val_loss"),
        "source_checkpoint": str(args.source),
    }

    args.output.parent.mkdir(parents=True, exist_ok=True)
    torch.save(inference_checkpoint, args.output)

    size_mib = args.output.stat().st_size / 1024 / 1024
    print(f"Wrote {args.output} ({size_mib:.1f} MiB)")
    print(f"Keys: {sorted(inference_checkpoint.keys())}")
    print(f"Val loss: {inference_checkpoint.get('val_loss')}")


if __name__ == "__main__":
    main()