What problem does it solve?
Kernel and numerical changes in the flash-linear-attention library can silently break supported configurations, change tolerances, or regress performance. This Skill enforces a contract-first design discipline before implementation so every reachable configuration is classified, budgeted, tested, and benchmarked.
Core Features & Use Cases
- Contract cell enumeration: Enumerate every reachable combination of operator, backend, leaf pipeline, chunk geometry, bound regime, routing capability, dtype staging, and input numerical domain, with each row ending in an optimized path, existing fallback, or explicit unsupported error.
- Numerical budgets per stage: Record load, operand, accumulation, store, range, and error budgets for each leaf stage in both forward and backward directions, with documented algebraic derivations for safety thresholds.
- Three-layer coverage model: Separate production-representative hard gates, public-contract boundary gates, and beyond-contract adversarial coverage, with route-parity tests when multiple routes claim the same cell.
- Use Case: Before adding a new TileLang backend for Gated DeltaNet, use this Skill to build the contract table, record per-stage numerical budgets, define fallback semantics, and freeze the benchmark baseline.
Quick Start
Ask the AI to apply the fla-design-coverage contract process to plan a new kernel backend or numerical change before writing any implementation code.
