---

name: block_fisher_rearrangement description: Rearrange binary pruning masks using block-diagonal Fisher interactions while preserving per-layer cardinality.

Block Fisher Rearrangement

Use this skill after diagonal Fisher mask search has produced binary layer masks. It improves pruning locations within each Transformer layer by using a block Fisher matrix. Do not use it to change the pruning ratio; the paper's rearrangement stage preserves per-layer kept counts and therefore preserves cost.

Inputs

• A binary mask for each layer.
• A square Fisher block for each layer.
• Optional maximum number of greedy passes.

Outputs

• Rearranged masks with the same cardinality as inputs.
• Objective values before and after rearrangement.
• Swap trace.

Workflow

For each layer, score (1-m)^T I (1-m). Try swaps between pruned and kept units and commit only objective-improving swaps. Repeat until no improvement or a pass limit is reached.

Validation

Run the included tests or validate_skill_tree.py --run-tests.