robottt-pi05-base-comparison / code /scripts /prepare_legacy_l18_protocol.py
jisue's picture
Experiment backup 2026-11-11T06:35:51.798729+00:00
fffb321 verified
Raw History Blame Contribute Delete
4.69 kB
"""Extend the actual L4 initial state to L18 and retain its exact training inputs."""
import json
import os
from pathlib import Path
import subprocess
import sys
import numpy as np
import torch
ROOT = Path(__file__).resolve().parents[1]
LOCAL = ROOT / '.release_local'
sys.path.insert(0, str(ROOT / 'src'))
from robottt_pi05.config import ExperimentConfig
from robottt_pi05.core import TTTLayer, TTTConfig
from robottt_pi05.modeling import extension
from robottt_pi05.eval_io import atomic_json, atomic_tensor, sha, parameter_hash
def main():
bundle = LOCAL / 'legacy-l4-handoff/bundle'
original = bundle / 'learned_inner_lr_v1/protocol/initial.pt'
assert sha(original) == '2e59f08afdd6bc863ffc89ecd29ff7ba25e861bdccfba6ae44f9b63eb210830b'
saved = torch.load(original, map_location='cpu', weights_only=True)
cfg_path = ROOT / 'configs/l18_learned_s4.json'
config = ExperimentConfig.load(cfg_path)
slow = extension(config)
shared = {'4', '8', '12', '17'}
seeds = {}
for key, layer in slow.ttt.items():
if key not in shared:
seed = int(np.random.SeedSequence([281004, int(key)]).generate_state(1, dtype=np.uint32)[0])
seeds[key] = seed
with torch.random.fork_rng(devices=[]):
torch.manual_seed(seed)
initialized = TTTLayer(TTTConfig(dim=1024, memory_dim=32, hidden_dim=64, base_lr=1e-4))
layer.load_state_dict(initialized.state_dict(), strict=True)
state = slow.state_dict()
for key, value in saved['slow'].items():
state[key].copy_(value)
for key, value in saved['slow'].items():
assert torch.equal(state[key], value), key
assert len({p.data_ptr() for p in slow.parameters()}) == len(list(slow.parameters()))
protocol = LOCAL / 'protocol-legacy-ft-l18'
protocol.mkdir(exist_ok=True)
initial = protocol / 'initial.pt'
initial_record = dict(slow=state, rng=saved['rng'], parameter_names=[n for n, p in slow.named_parameters() if p.requires_grad],
source_initial=str(original), source_step0=str(original), source_step0_sha256=sha(original), not_smoke=True)
atomic_tensor(initial, initial_record)
for filename, source_name in [('schedule.json', 'schedule.json'), ('bank.pt', 'banks.pt')]:
source = bundle / 'sequence_training_v1/protocol' / source_name
target = protocol / filename
if not target.exists():
os.link(source, target)
data = LOCAL / 'legacy-ft-data'
index = LOCAL / 'legacy-ft-prefix/asset-index.json'
paths = dict(openpi_source=str(ROOT.parent / 'openpi'),
model_checkpoint=str(LOCAL / 'ft-base-reproduction/converted/pi05_libero_pytorch/model.safetensors'),
normalization=str(data / 'normalization.json'), tokenizer=str(LOCAL / 'assets/paligemma_tokenizer.model'),
data_root=str(data), prefix_root=str(LOCAL / 'legacy-ft-prefix'), asset_index=str(index),
initial_extension=str(initial), schedule=str(protocol / 'schedule.json'), bank=str(protocol / 'bank.pt'),
output=str(LOCAL / 'runs/legacy-ft-l18-learned'), experiment_config=str(cfg_path))
hashes = {name: sha(Path(paths[name])) for name in ('model_checkpoint', 'normalization', 'tokenizer', 'asset_index', 'initial_extension', 'schedule', 'bank')}
uuid = subprocess.check_output(['nvidia-smi', '--query-gpu=uuid', '--format=csv,noheader'], text=True).strip()
run = dict(paths=paths, hashes=hashes, gpu_uuid=uuid, updates=1024, milestones=[256, 512, 1024], checkpoint_every=25,
runtime_profile='minimal', model_profile='legacy_pi05_libero', optimizer_profile='legacy', original=False,
limits=dict(model_load=4, expert=16384, inner=294912, backward=4096, optimizer_update=1024, rollout=0))
atomic_json(LOCAL / 'legacy-ft-l18-learned.json', run)
atomic_json(protocol / 'initialization-audit.json', dict(shared_tensors_verified=len(saved['slow']), shared_layers=[4,8,12,17],
register_initial_exact=True, original_initial_sha256=sha(original), added_layer_seeds=seeds,
trainable_parameters=sum(p.numel() for p in slow.parameters() if p.requires_grad),
trainable_tensors=sum(p.requires_grad for p in slow.parameters()), initial_parameter_hash=parameter_hash(state),
optimizer='legacy AdamW 3e-4, betas .9/.999, clip 1, fixed outer LR',
runtime='sequential S4 graph with reduced diagnostic bookkeeping; no timestep batching'))
print('L18_PROTOCOL_READY', str(LOCAL / 'legacy-ft-l18-learned.json'), flush=True)
if __name__ == '__main__':
main()