"""Commands are explicit; no training or model load on package import.""" import argparse,json def main(): parser=argparse.ArgumentParser(prog='robottt-pi05') sub=parser.add_subparsers(dest='command',required=True) sub.add_parser('quickstart',help='CPU synthetic inner-update demo; no optimizer/model assets') prep=sub.add_parser('prepare-protocol',help='Generate new seeded initialization, sequence schedule and noise bank') prep.add_argument('--experiment-config',required=True) prep.add_argument('--asset-index',required=True) prep.add_argument('--output',required=True) prep.add_argument('--updates',required=True,type=int) prep.add_argument('--seed',default=1000,type=int) reg=sub.add_parser('regression',help='Optional four saved-reference chunks; explicit assets required') reg.add_argument('--manifest',required=True) for command in ('train','evaluate'): p=sub.add_parser(command);p.add_argument('--run-config',required=True) if command=='train': p.add_argument('--resume',action='store_true') p.add_argument('--stop-after',type=int,help='Exit after this many additional updates at a saved boundary') args=parser.parse_args() import torch torch.set_num_threads(2) if args.command=='prepare-protocol': from .prepare import prepare_protocol print(json.dumps(prepare_protocol(config_path=args.experiment_config,asset_index=args.asset_index,output=args.output,updates=args.updates,seed=args.seed),indent=2));return if args.command=='quickstart': from .core import TTTLayer,TTTConfig with torch.random.fork_rng(devices=[]): torch.manual_seed(7);layer=TTTLayer(TTTConfig(base_lr=1e-4)).double() x=torch.arange(12,dtype=torch.float64).reshape(1,3,4)/20 y,state=layer.step(x,layer.initial_state(1),differentiable=False) print(json.dumps({'synthetic_only':True,'output_shape':list(y.shape),'counter':state.updates.tolist(),'optimizer_updates':0,'VLA_loads':0})) return if args.command=='regression': from .regression import run print(json.dumps(run(args.manifest),indent=2));return from .runner import train,evaluate result=train(args.run_config,resume=args.resume,stop_after=args.stop_after) if args.command=='train' else evaluate(args.run_config) print(json.dumps(result,indent=2))