import sys from pathlib import Path project_root = Path(__file__).parent.parent sys.path.append(str(project_root)) if __name__ == "__main__": # Keep CUDA/model imports out of multiprocessing worker module imports. # Only the torchrun training ranks execute these registrations. import thinkstream.model.patch # noqa: F401 from thinkstream.model.streaming_attention import register_streaming_attention from thinkstream.trainer.builder import TRAINER_BUILDERS register_streaming_attention() if len(sys.argv) < 2: print("Usage: torchrun ... train.py [args...]") print(f"Available builders: {list(TRAINER_BUILDERS.keys())}") sys.exit(1) builder_name = sys.argv[1] try: # Retrieve the builder from the registry builder_fn = TRAINER_BUILDERS.get(builder_name) except KeyError: print(f"Error: Unknown builder '{builder_name}'.") print(f"Available builders: {list(TRAINER_BUILDERS.keys())}") sys.exit(1) # Build the train node train_node = builder_fn() # Parse CLI arguments, validate Arg inputs, prepare, and execute the pipeline. train_node.run( use_argparse=True, cli_args=sys.argv[2:], ) print(train_node)