File size: 5,279 Bytes
36b714d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | """Rank-8 policy using explicitly assumed official RoboTwin preprocessing."""
import hashlib
import json
from pathlib import Path
ADAPTER_SHA = '51563f4e588174361ef8ad61db63a23da4ec33c0b115095375bd9323025cf5f0'
CONFIG_SHA = 'f0c1f8f2e9add85ac9bf573feba91ce52fa4bc24fb4927f175a8b7520e877449'
REVISION = '1d40e50988de2da6a07877e81a0941da3317733b'
MODEL = 'lora_r8_step10000'
def sha(path):
h = hashlib.sha256()
with Path(path).open('rb') as f:
for data in iter(lambda: f.read(8 * 1024 * 1024), b''):
h.update(data)
return h.hexdigest()
def load_lora_policy(root, output):
import torch
import yaml
from peft import LoraConfig
from safetensors import safe_open
from safetensors.torch import load_file
from src.phase4_policy import prepare_base_runtime, streaming_server_class
from src.lora_eval_loading import expand_fused_experts, inject_and_load_adapter
root, output = Path(root), Path(output)
adapter = root / 'outputs/lora_r8_smoke_preparation_20260929/adapter'
if sha(adapter/'adapter_model.safetensors') != ADAPTER_SHA or sha(adapter/'adapter_config.json') != CONFIG_SHA:
raise ValueError('Frozen adapter checksum mismatch')
upstream = root/'third_party/lingbot-vla-v2'
norm = upstream/'assets/norm_stats/robotwin.json'
base = root/'checkpoints/base/lingbot-vla-v2-6b'
# Verify the complete official base, not merely its model name or shapes.
manifest = json.loads((root/'manifests/lingbot-vla-v2-6b.json').read_text())
if manifest['revision'] != '11c703bf6a5c1f45b3b69168482da11fdbba53d7':
raise ValueError('Wrong base revision')
index = json.loads((base/'model.safetensors.index.json').read_text())
siblings = {x['rfilename']: x for x in manifest['siblings']}
for name in sorted(set(index['weight_map'].values())):
if sha(base/name) != siblings[name]['lfs']['sha256']:
raise ValueError('Base shard checksum mismatch: ' + name)
checkpoint, provenance = prepare_base_runtime(base,
root/'checkpoints/dependencies/Qwen3-VL-4B-Instruct', norm,
output/'runtime', upstream=upstream)
yaml_path = output/'runtime/lingbotvla_cli.yaml'
cfg = yaml.safe_load(yaml_path.read_text())
cfg['model']['moe_implementation'] = 'eager'
yaml_path.write_text(yaml.safe_dump(cfg, sort_keys=False))
class LoraServer(streaming_server_class(upstream)):
def load_model_weights(self, path_to_pi_model, strict=True):
if not strict:
raise ValueError('Strict base loading required')
expected = self.vla.state_dict()
tensors = {}
for name in sorted(set(index['weight_map'].values())):
with safe_open(base/name, framework='pt', device='cpu') as shard:
raw = {}
for key in shard.keys():
if index['weight_map'].get(key) != name:
raise ValueError('Wrong shard assignment: ' + key)
t = shard.get_tensor(key)
raw[key] = t.float() if t.is_floating_point() else t
for key, t in expand_fused_experts(raw).items():
if key in tensors or key not in expected or t.shape != expected[key].shape:
raise ValueError('Converted base key/shape mismatch: ' + key)
tensors[key] = t
print('Loaded and expanded base shard', name, flush=True)
if set(tensors) != set(expected):
raise ValueError('Converted base inventory does not match eager policy')
self.vla.load_state_dict(tensors, strict=True, assign=True)
if any(t.is_meta for t in list(self.vla.parameters()) + list(self.vla.buffers())):
raise ValueError('Meta tensors remain after loading')
self.vla = inject_and_load_adapter(self.vla,
LoraConfig.from_pretrained(str(adapter)),
load_file(str(adapter/'adapter_model.safetensors'), device='cpu'))
self.loaded_parameter_count = sum(p.numel() for p in self.vla.parameters())
self.loaded_tensor_count = len(self.vla.state_dict())
print('Strict eager base + complete LoRA loaded', self.loaded_parameter_count, flush=True)
policy = LoraServer(path_to_pi_model=str(checkpoint), robot_norm_path=str(norm),
use_bf16=False, use_fp32=True, use_compile=False, use_length=50, chunk_ret=True)
policy.reset('robotwin')
metadata = {'checkpoint_source_kind': 'lora_adapter', 'adapter_revision': REVISION,
'adapter_sha256': ADAPTER_SHA, 'adapter_config_sha256': CONFIG_SHA,
'base_revision': manifest['revision'], 'base_shards_verified': True,
'adapter_tensors_verified': 8494, 'lora_rank': 8, 'lora_alpha': 8,
'moe_implementation': 'eager', 'training_configuration_verified': False,
'configuration_assumption': 'User-authorized official RoboTwin model/preprocessing/normalization; eager expert layout for LoRA',
'norm_sha256': sha(norm), 'runtime_yaml_sha256': sha(yaml_path)}
(output/'checkpoint_verified.json').write_text(json.dumps(metadata, indent=2)+'\n')
return policy, norm, metadata
|