What problem does it solve? Designing thorough tests for TileLang-Ascend NPU operators is error-prone: teams often miss non-aligned tail shapes, dtype combinations, or boundary values, and lack a machine-checkable way to prove coverage. This Skill turns operator design docs or kernel code into a complete, layered test plan with enforced coverage gates. ## Core Features & Use Cases - Four input scenarios: generate tests from design.md, supplement tests for existing custom/{op}/.py kernels, build templates from a verbal operator description, or analyze existing test coverage for gaps. - Four-layer test system: L0 smoke tests, L1 functional tests with deterministically generated non-aligned/prime shapes, L2 negative exception tests, and Boundary tests for INF/NAN/extreme values. - Contract-based coverage gate: every L1 case carries D- dimension tags, and scripts/coverage_check.py validates the test file against the coverage matrix, blocking delivery on any unexempted MISS. - Use Case: Given a fused BatchMatmulMaxSum operator design doc, the Skill classifies it as a Cube+Vector fusion operator, generates dtype/shape/value-range test cases with mixed-tolerance precision standards per dtype, and verifies coverage before sign-off. ## Quick Start Ask the assistant to design a layered test plan for your TileLang-Ascend operator by pointing it to your design.md or existing kernel file under custom/{op}/.
