Normalize only spatial predecessor phases; remove NumPy usage; preserve CPU statistics configuration; lazy-load Lightning exports

[stephandooper]

Motivation

• Fix incorrect normalization in _compatible_predecessor_coordinates that applied remainder across the legacy (non-spatial) coordinate and could produce zero-divisor or lattice-mismatch issues for spatial branches.
• Remove an undeclared runtime dependency on NumPy in the max-pool gradient reconstruction path and ensure CPU-statistics configuration is preserved when gathering statistics.
• Allow core imports to succeed when optional runtime packages such as lightning are not installed by making convenience exports lazy.

Description

• Change the predecessor collector traversal to collect every nearest statistics-bearing spatial predecessor and deduplicate results, implemented in _prev_stats to return a list of stats rather than a single entry and to stop traversal per-path. (_prev_stats, _compatible_predecessor_coordinates)
• Restrict remainder normalization and stride/phase comparisons to height/width coordinates only, validate that spatial effective strides are strictly positive, and normalize phases modulo the spatial stride. (_compatible_predecessor_coordinates)
• Replace NumPy-based repeats/allocation with pure PyTorch operations in the max-pool gradient reconstruction (torch.repeat_interleave and torch.zeros) to remove the numpy runtime dependency and keep tensors on the correct device.
• Preserve the model’s original device when statistics_on_cpu is used by recording original_device and restoring it after CPU statistics gathering instead of unconditionally moving the model to CUDA. (_configure)
• Make LightningStreamingModule and ImageNetClassifier lazy imports via __getattr__ to avoid import-time failures when optional lightning/torchvision packages are not installed. (lightstream/modules/__init__.py)

Testing

• Ran the focused predecessor tests with pytest -q tests/test_scnn.py -k 'prev_stats or predecessor_coordinates' and python -m compileall -q lightstream, which produced 7 passed and successful byte-compilation respectively, confirming the zero-divisor / remainder cases are resolved.

• Ran the full test file with pytest -q tests/test_scnn.py, which produced 57 passed, 10 failed, 1 warning; none of the failures point to the predecessor collector including parameter/non-spatial branches or incorrectly rejecting valid legacy graphs, and stride metadata computed during statistics collection appears preserved.

• Remaining failures (grouped by exception and originating test locations) are:

  • AssertionError at tests/test_scnn.py:694, tests/test_scnn.py:850, and tests/test_scnn.py:1074 related to GeM saliency parity, mixed head mapping parity, and the reducer mask error-message contract.
  • ValueError at tests/test_scnn.py:866, tests/test_scnn.py:888, and tests/test_scnn.py:1085 originating from lightstream/core/reducer/base.py:420 where converted reducers enter a legacy passthrough expecting a single tensor (multi-input reducer handling issue).
  • RuntimeError at tests/test_scnn.py:743 indicating a backward replay / grad_fn mismatch for shared-parameter GeM reducers.
  • AttributeError at tests/test_scnn.py:129 related to a test constructing StreamingCNN with __new__ then assigning modules before Module.__init__ is called (test fixture limitation).
  • ModuleNotFoundError at tests/test_scnn.py:1193 due to the test requiring optional torchvision in the environment.

• Conclusion: the spatial-only remainder fix resolves the zero-divisor / lattice-normalization problems and the collector behaves as intended; the remaining failures point to reducer conversion/passthrough logic, backward-replay handling for shared parameters, and a couple of environment/fixture issues rather than regressions in spatial coordinate collection or stride metadata.

---

Codex Task