File size: 11,622 Bytes
e588892 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 | #!/usr/bin/env python
"""Run a bounded ValueCriticModel data/forward/backward/checkpoint smoke.
This deliberately uses one local RECAP episode and the tiny ``gemma_1m`` value
expert. It validates the wiring and memory envelope; it is not a useful value
model training run and its checkpoint must not be used for a paper result.
"""
from __future__ import annotations
import argparse
import gc
import json
import logging
import os
from pathlib import Path
os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
import torch
from omegaconf import OmegaConf
from rlinf.data.datasets.recap.value_model import ValueDataset
from rlinf.models.embodiment.value_model import get_model
from rlinf.models.embodiment.value_model.data_collator import ValueDataCollator
from rlinf.models.embodiment.value_model.processing import ValueProcessor
def _move(value, device):
if isinstance(value, torch.Tensor):
return value.to(device)
if isinstance(value, dict):
return {k: _move(v, device) for k, v in value.items()}
return value
def _grad_summary(model: torch.nn.Module) -> tuple[int, float]:
count = 0
squared = 0.0
for param in model.parameters():
if param.grad is None:
continue
count += 1
squared += float(param.grad.detach().float().pow(2).sum().item())
return count, squared**0.5
def parse_args() -> argparse.Namespace:
root = Path(os.environ.get("PROJECT_ROOT", "/data/myx/RL_post_train"))
parser = argparse.ArgumentParser()
parser.add_argument(
"--dataset",
type=Path,
default=root / "datasets/recap_value_smoke/libero10_task0_train",
)
parser.add_argument(
"--siglip", type=Path, default=root / "checkpoints/pretrained/siglip2-so400m-patch14-224"
)
parser.add_argument(
"--gemma3", type=Path, default=root / "checkpoints/pretrained/gemma-3-270m"
)
parser.add_argument(
"--tokenizer",
type=Path,
default=None,
help="Tokenizer directory; defaults to --gemma3 (use SigLIP tokenizer for the random smoke fallback).",
)
parser.add_argument(
"--output",
type=Path,
default=root / "checkpoints/value/value_smoke_gemma1m",
)
parser.add_argument("--critic-expert-variant", default="gemma_1m")
parser.add_argument("--max-token-len", type=int, default=64)
parser.add_argument("--freeze-vlm", action="store_true")
parser.add_argument("--freeze-vision", action="store_true")
parser.add_argument("--skip-backward", action="store_true")
parser.add_argument(
"--steps",
type=int,
default=10,
help="Number of repeated steps on the same micro-batch (wiring smoke only).",
)
parser.add_argument("--device", default="cuda")
return parser.parse_args()
def main() -> int:
args = parse_args()
if args.steps < 1:
raise ValueError("--steps must be >= 1")
logging.basicConfig(level=logging.INFO)
for name, path in (("dataset", args.dataset), ("siglip", args.siglip), ("gemma3", args.gemma3)):
if not path.exists():
raise FileNotFoundError(f"{name} path does not exist: {path}")
if not torch.cuda.is_available() and args.device.startswith("cuda"):
raise RuntimeError("CUDA is required for this preflight; torch.cuda.is_available() is false")
device = torch.device(args.device)
torch.manual_seed(0)
if device.type == "cuda":
torch.cuda.manual_seed_all(0)
torch.set_float32_matmul_precision("high")
dataset = ValueDataset(
dataset_path=str(args.dataset),
robot_type="libero",
model_type="pi05",
action_horizon=10,
action_dim=7,
return_min=-779.0,
return_max=0.0,
normalize_to_minus_one_zero=True,
tag="formula_v1",
)
processor = ValueProcessor(
max_token_len=args.max_token_len,
tokenizer_name_or_path=str(args.tokenizer or args.gemma3),
do_augment=False,
)
collator = ValueDataCollator(
processor=processor,
max_length=args.max_token_len,
train=False,
)
batch = collator([dataset[0]])
observation = _move(batch["observation"], device)
target_values = _move(batch["target_values"], device)
print(
"batch=",
json.dumps(
{
"images": {k: list(v.shape) for k, v in observation["images"].items()},
"image_masks": {k: list(v.shape) for k, v in observation["image_masks"].items()},
"tokens": list(observation["tokenized_prompt"].shape),
"target_values": target_values.detach().cpu().tolist(),
},
sort_keys=True,
),
flush=True,
)
cfg = OmegaConf.create(
{
"precision": "bf16",
"critic_expert_variant": args.critic_expert_variant,
"num_bins": 201,
"v_min": -1.0,
"v_max": 0.0,
"value_dropout": 0.0,
"action_dim": 7,
"action_horizon": 10,
"max_token_len": args.max_token_len,
"siglip_path": str(args.siglip),
"gemma3_path": str(args.gemma3),
"freeze_vlm": bool(args.freeze_vlm),
"freeze_vision_encoder": bool(args.freeze_vision),
"train_expert_only": bool(args.freeze_vlm),
"forward_mode": "vla",
"stop_gradient_to_vlm": bool(args.freeze_vlm),
}
)
print(
"config=",
json.dumps(OmegaConf.to_container(cfg, resolve=True), sort_keys=True),
flush=True,
)
print("constructing model", flush=True)
model = get_model(cfg)
total_params = sum(p.numel() for p in model.parameters())
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(
f"params total={total_params} trainable={trainable_params} "
f"cpu_bytes~={sum(p.numel() * p.element_size() for p in model.parameters())}",
flush=True,
)
model = model.to(device)
model.train(not args.skip_backward)
if device.type == "cuda":
torch.cuda.reset_peak_memory_stats(device)
reference_prediction = None
optimizer = None
if not args.skip_backward:
optimizer = torch.optim.AdamW(
[p for p in model.parameters() if p.requires_grad], lr=1e-4
)
last_grad_count = 0
last_grad_norm = 0.0
print(f"pipeline_steps={args.steps}", flush=True)
for step in range(args.steps):
if optimizer is not None:
optimizer.zero_grad(set_to_none=True)
output = model(observation=observation, target_values=target_values)
if output.loss is None or not torch.isfinite(output.loss).all():
raise RuntimeError(f"non-finite value loss at step {step}: {output.loss}")
if not torch.isfinite(output.predicted_values).all():
raise RuntimeError(
f"non-finite predicted values at step {step}: {output.predicted_values}"
)
if optimizer is not None:
output.loss.backward()
last_grad_count, last_grad_norm = _grad_summary(model)
if last_grad_count == 0 or not torch.isfinite(torch.tensor(last_grad_norm)):
raise RuntimeError(
f"no finite gradients at step {step}: "
f"count={last_grad_count}, norm={last_grad_norm}"
)
optimizer.step()
if step in (0, args.steps - 1):
print(
"step_result=",
json.dumps(
{
"step": step + 1,
"loss": float(output.loss.detach().float().cpu()),
"predicted_values": output.predicted_values.detach()
.float()
.cpu()
.tolist(),
"logits_shape": list(output.logits.shape),
"mae": None
if output.mae is None
else float(output.mae.detach().float().cpu()),
},
sort_keys=True,
),
flush=True,
)
if optimizer is not None:
print(
f"backward_result=grad_tensors:{last_grad_count},"
f"grad_norm:{last_grad_norm:.6f}",
flush=True,
)
# Evaluate once after the update. The reload check below compares
# against this post-update value (rather than the pre-update forward
# value), making checkpoint restoration an actual numerical check.
model.eval()
with torch.no_grad():
reference_prediction = model(
observation=observation, target_values=None
).predicted_values.detach().cpu()
model.train()
# Do not retain the last autograd graph while constructing the fresh model
# below; keeping it alive would unnecessarily raise the reload peak.
if "output" in locals():
del output
args.output.mkdir(parents=True, exist_ok=True)
state_dir = args.output / "model_state_dict"
state_dir.mkdir(parents=True, exist_ok=True)
full_weights = state_dir / "full_weights.pt"
state_cpu = {k: v.detach().cpu() for k, v in model.state_dict().items()}
torch.save(state_cpu, full_weights)
(args.output / "config.json").write_text(
json.dumps(OmegaConf.to_container(cfg, resolve=True), indent=2, sort_keys=True) + "\n"
)
print(f"saved_checkpoint={full_weights} bytes={full_weights.stat().st_size}", flush=True)
# A fresh model verifies that the exact checkpoint format can be loaded.
if device.type == "cuda":
model = model.cpu()
torch.cuda.empty_cache()
if not args.skip_backward:
del optimizer
del model
gc.collect()
if device.type == "cuda":
torch.cuda.empty_cache()
print("reloading model", flush=True)
reloaded = get_model(cfg)
missing, unexpected = reloaded.load_state_dict(torch.load(full_weights, map_location="cpu", weights_only=False), strict=False)
if missing or unexpected:
raise RuntimeError(f"checkpoint mismatch: missing={missing}, unexpected={unexpected}")
reloaded = reloaded.to(device).eval()
with torch.no_grad():
reloaded_output = reloaded(observation=observation, target_values=None)
if not torch.isfinite(reloaded_output.predicted_values).all():
raise RuntimeError("reloaded model produced non-finite prediction")
if reference_prediction is not None:
reloaded_cpu = reloaded_output.predicted_values.detach().float().cpu()
if not torch.allclose(reloaded_cpu, reference_prediction.float(), atol=1e-5, rtol=1e-4):
raise RuntimeError(
"checkpoint reload prediction differs from post-update reference: "
f"reference={reference_prediction.tolist()} reload={reloaded_cpu.tolist()}"
)
print(
"reload_result=",
json.dumps(
{"predicted_values": reloaded_output.predicted_values.detach().float().cpu().tolist()},
sort_keys=True,
),
flush=True,
)
if device.type == "cuda":
print(
f"cuda_peak_allocated={torch.cuda.max_memory_allocated(device)} "
f"cuda_peak_reserved={torch.cuda.max_memory_reserved(device)}",
flush=True,
)
print("VALUE_PREFLIGHT_PASS", flush=True)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|