File size: 9,544 Bytes
6f003f2 | 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 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 | #!/usr/bin/env python
"""Single-card CFG model load/forward preflight for RLinf's π0.5 path.
This is deliberately bounded to one transformed fixture batch. It does not
perform an optimizer step and therefore cannot establish training viability.
"""
from __future__ import annotations
import argparse
import json
import logging
from pathlib import Path
import numpy as np
import torch
from omegaconf import OmegaConf
from torch.utils.data import DataLoader
import lerobot.common.datasets.lerobot_dataset as lerobot_dataset
import openpi.training.data_loader as openpi_data_loader
import openpi.transforms as transforms
from rlinf.data.datasets.recap.cfg_model import (
AdvantagePreservingDataset,
TokenizePromptWithGuidance,
)
from rlinf.data.datasets.recap.utils import cast_image_features
from rlinf.data.lerobot_paths import resolve_lerobot_dataset_root
from rlinf.models.embodiment.openpi.dataconfig import get_openpi_config
class _LookupLimitedDataset:
def __init__(self, delegate, count: int):
self._delegate = delegate
self.hf_dataset = delegate.hf_dataset.select(range(count))
def __len__(self):
return len(self.hf_dataset)
def __getitem__(self, index):
return self._delegate[index]
def _move(value, device):
if isinstance(value, torch.Tensor):
return value.to(device, non_blocking=True)
if isinstance(value, dict):
return {key: _move(item, device) for key, item in value.items()}
return value
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser()
parser.add_argument(
"--dataset",
type=Path,
default=Path(
"/data/myx/RL_post_train/datasets/recap_value_smoke/"
"libero10_task0_train"
),
)
parser.add_argument(
"--model",
type=Path,
default=Path(
"/data/myx/RL_post_train/checkpoints/base/RLinf-Pi05-LIBERO-SFT"
),
)
parser.add_argument("--advantage-tag", default="smoke")
parser.add_argument("--max-samples", type=int, default=8)
parser.add_argument("--device", default="cuda")
return parser.parse_args()
def _build_batch(dataset_path: Path, model_path: Path, advantage_tag: str, max_samples: int):
data_path = str(dataset_path.resolve())
train_config = get_openpi_config(
"pi05_libero", model_path=str(model_path.resolve()), batch_size=1, repo_id=data_path
)
data_config = train_config.data.create(train_config.assets_dirs, train_config.model)
dataset_root = resolve_lerobot_dataset_root(data_path)
dataset_meta = lerobot_dataset.LeRobotDatasetMetadata(data_path, root=dataset_root)
base_dataset = lerobot_dataset.LeRobotDataset(
data_path,
root=dataset_root,
delta_timestamps={
key: [t / dataset_meta.fps for t in range(train_config.model.action_horizon)]
for key in data_config.action_sequence_keys
},
)
base_dataset.hf_dataset = cast_image_features(base_dataset.hf_dataset)
base_dataset = _LookupLimitedDataset(
base_dataset, min(max_samples, len(base_dataset.hf_dataset))
)
if data_config.prompt_from_task:
base_dataset = openpi_data_loader.TransformedDataset(
base_dataset,
[transforms.PromptFromLeRobotTask(dataset_meta.tasks)],
)
tokenizer = None
for transform in data_config.model_transforms.inputs:
if hasattr(transform, "tokenizer"):
tokenizer = transform.tokenizer
break
if tokenizer is None:
raise RuntimeError("Could not find tokenizer in pi05_libero transforms")
model_transforms = []
for transform in data_config.model_transforms.inputs:
if type(transform).__name__ == "TokenizePrompt":
model_transforms.append(
TokenizePromptWithGuidance(
tokenizer=tokenizer,
discrete_state_input=getattr(transform, "discrete_state_input", False),
)
)
else:
model_transforms.append(transform)
transformed_dataset = openpi_data_loader.TransformedDataset(
base_dataset,
[
*data_config.repack_transforms.inputs,
*data_config.data_transforms.inputs,
transforms.Normalize(
data_config.norm_stats, use_quantiles=data_config.use_quantile_norm
),
*model_transforms,
],
)
adv_path = dataset_path / "meta" / f"advantages_{advantage_tag}.parquet"
adv_df = __import__("pandas").read_parquet(adv_path)
lookup = dict(
zip(
zip(adv_df["episode_index"].astype(int), adv_df["frame_index"].astype(int)),
adv_df["advantage"].astype(bool),
)
)
final_dataset = AdvantagePreservingDataset(
base_dataset=base_dataset,
transformed_dataset=transformed_dataset,
advantages_lookup=lookup,
)
batch = next(
iter(
DataLoader(
final_dataset,
batch_size=1,
shuffle=False,
num_workers=0,
drop_last=True,
)
)
)
return train_config, batch
def main() -> int:
args = parse_args()
logging.basicConfig(level=logging.INFO)
if not torch.cuda.is_available() and args.device.startswith("cuda"):
raise RuntimeError("CUDA is required for CFG model preflight")
device = torch.device(args.device)
train_config, batch = _build_batch(
args.dataset, args.model, args.advantage_tag, args.max_samples
)
# The parent PI0Pytorch constructor compiles sampling even though this
# preflight calls only forward. Eager mode keeps the test deterministic
# and avoids an unbounded compile warm-up.
original_compile = torch.compile
torch.compile = lambda fn, *unused_args, **unused_kwargs: fn # type: ignore[assignment]
try:
from rlinf.models import get_model
cfg = OmegaConf.create(
{
"model_type": "cfg_model",
"model_path": str(args.model.resolve()),
"precision": None,
"load_to_device": False,
"is_lora": False,
"lora_rank": 32,
"num_action_chunks": 10,
"action_dim": 7,
"use_proprio": True,
"num_steps": 5,
"add_value_head": False,
"openpi": {
"config_name": "pi05_libero",
"num_images_in_input": 2,
"noise_level": 0.5,
"action_chunk": 10,
"num_steps": 5,
"train_expert_only": False,
"action_env_dim": 7,
"noise_method": "flow_sde",
"add_value_head": False,
"value_after_vlm": False,
"value_vlm_mode": "mean_token",
"detach_critic_input": None,
"cfgrl_guidance_scale": 1.0,
"unconditional_prob": 0.1,
"guidance_type": "positive",
"positive_only_conditional": True,
},
}
)
print("config=", json.dumps(OmegaConf.to_container(cfg, resolve=True), sort_keys=True))
if device.type == "cuda":
torch.cuda.reset_peak_memory_stats(device)
model = get_model(cfg)
finally:
torch.compile = original_compile
if model is None:
raise RuntimeError("RLinf cfg_model registry returned None")
total_params = sum(parameter.numel() for parameter in model.parameters())
trainable_params = sum(
parameter.numel() for parameter in model.parameters() if parameter.requires_grad
)
model = model.to(device).eval()
batch = _move(batch, device)
from rlinf.models.embodiment.openpi_cfg.openpi_cfg_action_model import Observation
observation = Observation.from_dict(batch)
actions = batch["actions"].to(device=device, dtype=torch.float32)
advantage = batch["advantage"].to(device=device, dtype=torch.bool)
time = torch.full((actions.shape[0],), 0.5, dtype=torch.float32, device=device)
noise = torch.zeros_like(actions, dtype=torch.float32, device=device)
with torch.inference_mode():
loss, metrics = model(
data={"observation": observation, "actions": actions, "advantage": advantage},
time=time,
noise=noise,
)
if not torch.isfinite(loss).all():
raise RuntimeError(f"non-finite CFG loss: {loss}")
print(
"CFG_MODEL_PREFLIGHT_PASS",
json.dumps(
{
"total_params": total_params,
"trainable_params": trainable_params,
"loss": float(loss.detach().float().cpu()),
"actions_shape": list(actions.shape),
"advantage_shape": list(advantage.shape),
"metrics": {key: float(np.mean(value)) for key, value in metrics.items()},
"cuda_peak_allocated": None
if device.type != "cuda"
else torch.cuda.max_memory_allocated(device),
"cuda_peak_reserved": None
if device.type != "cuda"
else torch.cuda.max_memory_reserved(device),
},
sort_keys=True,
),
)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|