attention-variants-from-papers

1# Attention Variants from Papers
2 
3Use this skill when a paper changes how attention scores, branches, normalization, or head sharing work, but the module still needs to behave like a drop-in transformer attention block.
4 
5## Workflow
6 
71. Read the paper for invariants, not names.
8   Use [paper-to-implementation.md](references/paper-to-implementation.md) to extract the external contract, the changed computation, and the training-time constraints.
92. Build a shape ledger before coding.
10   Use [shape-ledger.md](references/shape-ledger.md) to track projections, head grouping, branch count, and output width.
113. Choose the mechanism pattern.
12   Use [mechanism-patterns.md](references/mechanism-patterns.md) for subtractive attention, branch mixing, learned gates, and extra normalization.
134. Preserve the module boundary.
14   Keep the same input and output shape, mask semantics, positional encoding flow, and cache behavior unless the task explicitly changes them.
155. Validate in layers.
16   Start with random-tensor smoke tests, then compare against a baseline attention path. Use [stability-and-validation.md](references/stability-and-validation.md).
176. Integrate into the stack last.
18   Swap the new module into one transformer block, verify the residual path, then roll it through the full model. Use [transformer-integration.md](references/transformer-integration.md).
19 
20## Checklist
21 
22- extract the paper's invariants before writing code
23…