#!/usr/bin/env python3 """Export only the generator from DCP for the existing single-GPU .pth runner. Run with the pinned PyTorch 2.7 environment. This uses CPU RAM and no GPU; roughly 38 GB are allocated for a BF16 generator. Optimizers and critic are excluded; the result is for inference or weights-only initialization. """ from pathlib import Path import argparse import torch import torch.distributed.checkpoint as dcp def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument('checkpoint', type=Path) parser.add_argument('output', type=Path) args = parser.parse_args() if args.output.exists(): parser.error('Output already exists.') reader = dcp.FileSystemReader(str(args.checkpoint / 'sharded')) metadata = reader.read_metadata() # DCP planner paths preserve the nested "generator" dictionary and the # model.* state dict keys. Do not strip "model." from the wrapper state. generator = {} for key, storage in metadata.state_dict_metadata.items(): if key.startswith('generator.') and hasattr(storage, 'size'): path = metadata.planner_data[key] generator[path[1]] = torch.empty(storage.size, dtype=storage.properties.dtype, device='cpu') if not generator: raise RuntimeError('No generator tensors found in DCP metadata.') state = {'generator': generator} dcp.load(state, storage_reader=reader) source = torch.load(args.checkpoint/'metadata.pth', map_location='cpu', weights_only=False) payload = {'generator': generator, 'completed_step': source.get('completed_step'), 'next_step': source.get('next_step'), 'source_checkpoint': str(args.checkpoint.resolve()), 'has_generator_optimizer': False, 'has_critic_optimizer': False} args.output.parent.mkdir(parents=True, exist_ok=True) tmp = args.output.with_name(args.output.name + '.partial') torch.save(payload, tmp) tmp.replace(args.output) print(f'Exported {len(generator)} tensors, {args.output.stat().st_size} bytes: {args.output}') if __name__ == '__main__': main()