"""Rebuild original batch-one LIBERO-FT prefixes from hash-matched inputs.""" import json from pathlib import Path import shutil import sys import numpy as np import torch ROOT = Path(__file__).resolve().parents[1] LOCAL = ROOT / '.release_local' sys.path[:0] = [str(ROOT / 'src'), str(ROOT.parent / 'openpi/src'), str(ROOT.parent / 'openpi/packages/openpi-client/src')] from robottt_pi05.openpi_backend import load_model from robottt_pi05.eval_io import atomic_json, atomic_tensor, sha, tensor_hash from openpi.models.model import Observation from openpi.models_pytorch.pi0_pytorch import make_att_2d_masks def main(): torch.set_num_threads(4) torch.backends.cuda.matmul.allow_tf32 = True torch.backends.cudnn.allow_tf32 = True data = LOCAL / 'legacy-ft-data' out = LOCAL / 'legacy-ft-prefix' out.mkdir(exist_ok=True) manifest = json.loads((data / 'asset-index.json').read_text()) checkpoint = LOCAL / 'ft-base-reproduction/converted/pi05_libero_pytorch/model.safetensors' model_hash = sha(checkpoint) if model_hash != 'caf7fce67043d34faddbb9e37ac6f6ae33e2940f2bac9273464ad509cb9e1e31': raise ValueError('Original FT checkpoint hash differs') model = load_model(source=ROOT.parent / 'openpi', checkpoint=checkpoint, profile='legacy_pi05_libero') cameras = ['base_0_rgb', 'left_wrist_0_rgb', 'right_wrist_0_rgb'] count = 0 expected = sum(len(r['manifest']['rows']) for r in manifest['episodes'].values()) assert expected == 521 for episode, record in manifest['episodes'].items(): folder = data / record['folder'] arrays = {k: np.load(folder / name, mmap_mode='r') for k, name in record['manifest']['fields'].items()} for index, row in enumerate(record['manifest']['rows']): key = f'e{int(episode):06d}_r{row:06d}' destination = out / (key + '.pt') metadata = out / (key + '.json') if metadata.exists(): rec = json.loads(metadata.read_text()) if sha(destination) != rec['sha256'] or rec['provenance']['model_checkpoint_sha256'] != model_hash: raise ValueError('Existing prefix differs') else: tensors = {k: torch.from_numpy(np.array(v[index:index+1])).cuda() for k, v in arrays.items()} provenance = dict(episode=int(episode), source_row=row, inputs={k: tensor_hash(v.cpu()) for k, v in tensors.items()}, model_checkpoint_sha256=model_hash, model_revision='215abfb217dbac7d5f1273282331b9b1866c0479', dtype='BF16 prefix', attention='eager', positions='padding cumsum-1', normal_observation=True, selection_sha256=manifest['protocol']['selection_sha256']) observation = Observation(images={c: tensors['image/'+c] for c in cameras}, image_masks={c: tensors['image_mask/'+c] for c in cameras}, state=tensors['state'], tokenized_prompt=tensors['tokenized_prompt'], tokenized_prompt_mask=tensors['tokenized_prompt_mask']) with torch.no_grad(): images, masks, tokens, token_mask, state = model._preprocess_observation(observation, train=False) embedding, pm, pa = model.embed_prefix(images, masks, tokens, token_mask) embedding = embedding.to(torch.bfloat16) _, cache = model.paligemma_with_expert(attention_mask=model._prepare_attention_masks_4d(make_att_2d_masks(pm, pa)), position_ids=pm.cumsum(1)-1, inputs_embeds=[embedding, None], use_cache=True) legacy = cache.to_legacy_cache() if hasattr(cache, 'to_legacy_cache') else tuple(cache) value = dict(prefix_embedding=embedding.cpu(), padding_mask=pm.cpu(), attention_blocks=pa.cpu(), state=state.cpu(), KV=tuple((k.cpu(), v.cpu()) for k, v in legacy), provenance=provenance) if count == 0: size = sum(t.numel()*t.element_size() for t in [value['prefix_embedding'], value['padding_mask'], value['attention_blocks'], value['state'], *[t for pair in value['KV'] for t in pair]]) if shutil.disk_usage(out).free - expected*size < 160*2**30: raise RuntimeError('Insufficient prefix storage headroom') print('PREFIX_SIZE_BYTES', size, 'ESTIMATED_TOTAL', expected*size, flush=True) digest = atomic_tensor(destination, value) rec = dict(path=destination.name, sha256=digest, provenance=provenance) atomic_json(metadata, rec) del tensors, observation, images, masks, tokens, token_mask, state, embedding, pm, pa, cache, legacy, value manifest['prefixes'][key] = rec count += 1 atomic_json(out / 'progress.json', dict(prefixes=count, total=expected), replace=True) print('PREFIX_EPISODE', episode, count, flush=True) atomic_json(out / 'asset-index.json', manifest) print('LEGACY_PREFIXES_COMPLETE', count, flush=True) if __name__ == '__main__': main()