fix(pruning): validate multidimensional tensor in channel-structured pruning (#107)

[rohith500]

Summary of Changes

Closes #107.

Per maintainer alignment with @u-simha in Issue #107, this PR validates that ChannelStructured magnitude pruning is applied to tensors with at least 2 dimensions, raising an informative ValueError for 1D or 0D tensors instead of crashing with ATen RuntimeError: selected index k out of range.

1. Multidimensional Validation in _compute_channel_mask:
   • In src/coreai_opt/pruning/spec/prune.py:

     if weight.ndim < 2:
         raise ValueError(
             f"Channel-structured pruning requires a tensor with at least 2 dimensions, "
             f"got shape {tuple(weight.shape)} with {weight.ndim} dims. "
             f"For 1D tensors, use Unstructured pruning instead."
         )

   • Adheres to PyTorch core convention (torch.nn.utils.prune._validate_structured_pruning), ensuring that the concept of "channels" is well-defined.
   • Clear error message guides users to Unstructured pruning for 1D parameters (e.g. bias or normalization weights).
2. Early Scheme Dispatch in compute_mask:
   • Dispatches ChannelStructured directly to _compute_channel_mask so scheme-specific structural constraints are checked even if the schedule's initial sparsity is 0.0.
   • For multidimensional tensors ($N \ge 2$), _compute_channel_mask continues to handle sparsity == 0.0 (num_prune == 0) and sparsity >= 1.0 directly.

---

Automated Tests Added

In tests/pruning/test_magnitude_pruner.py:

• test_channel_structured_1d_tensor_raises:
  • Parametrically verifies that applying ChannelStructured to a 1D tensor via _MagnitudePruneImpl.compute_mask raises ValueError matching "Channel-structured pruning requires a tensor with at least 2 dimensions" across axis=0, axis=-1, and multiple sparsity levels (including 0.0).
• test_channel_structured_0d_scalar_raises:
  • Verifies that applying ChannelStructured to a 0-D scalar raises ValueError.
• test_channel_structured_1d_parameter_model_raises:
  • Tests end-to-end model preparation (MagnitudePruner.prepare) when a 1D parameter (bias of nn.Linear) is configured with ChannelStructured, verifying it raises ValueError.
• test_unstructured_1d_tensor_supported:
  • Verifies that Unstructured pruning properly supports 1D parameters (bias), confirming the suggested alternative in the error message.

---

Verification

• pytest tests/pruning/test_magnitude_pruner.py: 46 passed, 1 xfailed in 0.31s.
• pytest tests/pruning/: 85 passed, 1 xfailed in 3m 03s.
• make check: All checks passed (26/26):
  • License header verification: Passed
  • Formatting & line-length (ruff, darker $\le$ 100): Passed
  • Linters (ruff-lint, checkmake, pymarkdown 0.9.40): Passed
  • Project rules (__all__, import aliasing): Passed
  • Towncrier changelog validation: Passed

---

Checklist

• Follows Conventional Commits formatting (fix(pruning): validate multidimensional tensor in channel-structured pruning (#107))
• Commit is cryptographically signed with SSH key (SHA256:0wr/nMrZodwxsOOHBGpcwRXPk8c3rePKishJLzmhj98)
• Includes Towncrier fragment (changelog.d/107.fixed)
• Passes make check without warnings or diff pollution
• Includes automated tests reproducing the bug and proving the fix

[rohith500]

Hi @u-simha,

Per your feedback and confirmation on #107, I've updated PR #108 to enforce multidimensional tensor validation:

1. Validation Check: In _compute_channel_mask, added if weight.ndim < 2: raise ValueError(...) requiring $\ge 2$ dimensions and directing users to Unstructured pruning for 1D tensors.
2. Early Dispatch: In compute_mask, ChannelStructured is dispatched to _compute_channel_mask so structural constraints are validated upfront even if schedule sparsity is 0.0.
3. Automated Tests: Updated tests/pruning/test_magnitude_pruner.py with tests verifying ValueError is raised for 1D and 0-D tensors with ChannelStructured across multiple axes and sparsities, and confirming that Unstructured() properly prunes 1D parameters.
4. Quality Gates: All 26 checks in make check pass cleanly, commit is cryptographically signed, and the Towncrier fragment is updated.

Ready for your review!