"""Six compute-node checks using only actual supplied Darcy samples.""" from pathlib import Path import torch from torch.utils.data import DataLoader, TensorDataset from common import (arguments, atomic_json, build_model, configure_runtime, decode, evaluate, load_config, load_data, normalizer_from_train, relative_l2) def main(): args = arguments(__doc__).parse_args() config = load_config(args.config) device = configure_runtime(config) data, _ = load_data(config) normalizer = normalizer_from_train(config, data["train_y"]) model = build_model(config, device) checks = {} x, y = data["train_x"][:32].to(device), data["train_y"][:32].to(device) prediction = decode(model(x), normalizer) assert prediction.shape == y.shape and torch.isfinite(prediction).all() checks["forward_pass"] = "passed" loss = relative_l2(prediction, y).square().mean() loss.backward() assert all(p.grad is not None and torch.isfinite(p.grad).all() for p in model.parameters() if p.requires_grad) checks["backward_pass"] = "passed" optimizer = torch.optim.AdamW(model.parameters(), lr=0.001, weight_decay=0.0001) losses = [] for bx, by in DataLoader(TensorDataset(data["train_x"], data["train_y"]), batch_size=32): optimizer.zero_grad(set_to_none=True) prediction = decode(model(bx.to(device)), normalizer) batch_loss = relative_l2(prediction, by.to(device)).square().mean() assert torch.isfinite(batch_loss) batch_loss.backward() optimizer.step() losses.append(batch_loss.item()) assert sum(losses[-8:]) / 8 < sum(losses[:8]) / 8 checks["train_loop_dry_run"] = "passed" metrics, errors, pred = evaluate(model, data["test_x"][:32], data["test_y"][:32], normalizer, 32, device, return_predictions=True) assert torch.isfinite(errors).all() checks["validation_loop_dry_run"] = "passed" # Independent manual metric computation on real fields; no synthetic data. manual = ((pred.double() - data["test_y"][:32].double()).flatten(1).square().sum(1).sqrt() / data["test_y"][:32].double().flatten(1).square().sum(1).sqrt()) assert torch.allclose(errors.double(), manual, rtol=1e-5, atol=1e-8) checks["domain_metric_dry_run"] = "passed" assert len(model.blocks) == 4 and model.projection.layers[0].out_channels == 64 for block in model.blocks: assert tuple(block.spectral.coordinates.shape) == (32, 17, 2) checks["config_consistency"] = "passed" path = Path(config["output"]["root"]) / config["output"]["results"] / "smoke/checks.json" atomic_json(path, {"status": "passed", "attempts": 1, "checks": checks, "used_original_data": True, "formal_epochs_consumed": 0, "loss_first8_mean": sum(losses[:8]) / 8, "loss_last8_mean": sum(losses[-8:]) / 8, "validation_metric": metrics, "peak_memory_bytes": torch.cuda.max_memory_allocated()}) print("SMOKE: all six checks passed; discarded smoke model; formal run starts from fresh seed", flush=True) if __name__ == "__main__": main()