Instruction file imported from santanikolaus/pino-lightning (.github/instructions/*.instructions.md). Copyright stays with the author.

PyTorch Lightning + Hydra project for neural operators (FNO) solving PDEs.

Conventions

• Flag deviations from these conventions when generating code

1. Two blank lines after import block, one blank line between import groups (stdlib → third-party → local)
2. 4 spaces indentation, no tabs
3. Type hints on all function signatures including return types
4. No else after a return — flatten control flow
5. Max 20 lines per function body, excluding docstring; split otherwise
6. f-strings only — no .format() or %s
7. snake_case functions/variables, PascalCase classes, UPPER_CASE module-level constants; no abbreviations except domain-standard ones (lr, cfg, n_train)
8. Comments explain why, never what — if the code needs a "what" comment, rewrite it
9. No commented-out code in commits

Technicals

• Normalizers and DataProcessor are stateful only during .fit()/.partial_fit(); runtime transforms must only read buffers, never write them
• New functionality requires a corresponding test; check coverage for src/datasets/ and src/models/

Tensor Conventions

• TensorDataset.__getitem__ returns {"x": ..., "y": ...}; Lightning steps must keep dict form, never unpack positionally
• Every sample follows (B, C, H, W): PT datasets unsqueeze a channel dim and keep spatial dims after it; synthetic dataset tests assert those shapes explicitly
• DarcyLitModule preprocesses every batch through DataProcessor before the model and postprocesses predictions before loss/logging — never bypass this path

Hydra Setup

• All configs are Hydra dataclasses registered via ConfigStore; no hardcoded values in __init__ signatures
• AppConfig nests {data, model, opt, patching, trainer, loss}; physical parameters (resolution, domain size) belong in DataConfig only