"""LPU Virtual Runtime - intelligent GPU acceleration layer for PyTorch workloads. Primary target: Super-Resolution training on Windows 11 with dual Tesla P100 (Pascal, sm_60, no Tensor Cores). Designed for any CUDA-capable NVIDIA GPU. Design rules enforced across this package: * No fabricated metrics. Anything not measured is labeled an estimate. * NumPy is never described as CUDA. CUDA execution = torch CUDA tensors only. * Per-GPU memory accounting. VRAM is never treated as a unified pool. * The target project's Python environment is never modified (read-only). * Every optimization is reversible and fails open to plain PyTorch behavior. Torch is imported LAZILY (see lpu.compat): importing `lpu` never loads torch, so the background controller keeps a tiny torch-free footprint. Get torch via `from lpu.compat import require_torch`. """ from .version import __version__, PACKAGE_NAME, COMPATIBILITY from .compat import optional_import, require_torch, TORCH_AVAILABLE __all__ = [ "__version__", "PACKAGE_NAME", "COMPATIBILITY", "optional_import", "require_torch", "TORCH_AVAILABLE", ] def training_session(model, train_loader=None, config=None, name: str = "train", criterion=None, optimizer=None, batch_size=None): """Convenience constructor for lpu.training.TrainingSession (V4 kwargs).""" from .training import TrainingSession return TrainingSession(model, train_loader=train_loader, config=config, name=name, criterion=criterion, optimizer=optimizer, batch_size=batch_size)