#!/usr/bin/env python3 """Export a training checkpoint as a compact inference checkpoint. Keeps only the model weights (dropping optimizer state and the training arguments) and stores them in float16. Evaluation runs the model in fp16 (`--fp16` casts the model with `model.half()`), so the exported checkpoint gives the same scores as the original one. Usage: python tools/export_checkpoint.py save/causalbind_sp_seed1/savedir/checkpoint_last.pt \ checkpoints/causalbind_sp.pt --variant sp """ import argparse import hashlib import torch def main(): parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) parser.add_argument("src") parser.add_argument("dst") parser.add_argument("--variant", required=True, help="variant name in scripts/variant_config.sh") args = parser.parse_args() state = torch.load(args.src, map_location="cpu", weights_only=False) model = { k: (v.half() if torch.is_floating_point(v) else v) for k, v in state["model"].items() } torch.save({"model": model, "variant": args.variant}, args.dst) n_params = sum(v.numel() for v in model.values()) with open(args.dst, "rb") as handle: sha = hashlib.sha256(handle.read()).hexdigest() print(f"{args.dst}: {n_params / 1e6:.1f}M parameters, sha256 {sha}") if __name__ == "__main__": main()