ascend-triton-kernel-validation git:20260911.9f92063A

Immutable. This exact content is served forever at /api/v1/blob/f51189f1386137ce.

---
name: ascend-triton-kernel-validation
description: Validate one Ascend Triton kernel against a trusted reference across an explicit shape, dtype, layout, stride, scalar-option, and execution-mode case matrix, with static detection of missing kernel launches and PyTorch computation fallback plus Run Manifest evidence. Use before any performance claim, after migration or implementation changes, or for shape-dependent compile/runtime/numerical failures in a Triton candidate. Do not use to generate the kernel, optimize an already-correct kernel, diagnose a non-Triton torch_npu or ACLNN call, or localize a whole-model graph failure.
---

# ascend-triton-kernel-validation

Validate one Ascend Triton kernel against its reference over the cases needed by its consumers.

Select shapes, dtype, layout, strides, scalar options and execution modes from the operator contract and affected callers. Include boundary and non-contiguous cases where semantics require them. Numerical agreement must come from the launched candidate kernel.

## Agent entry

Run from the repository root using the platform's Python launcher. The workspace
selects its installed platform environment automatically.

```text
python .agents/skills/ascend-triton-kernel-validation/scripts/triton_validation.py --config validation.json --kernel kernel.py --results case-results.json
```

The config contains op_name, reference, target, cases and tolerances. The tool checks kernel source for missing launches and computation fallback, combines observed case results, and emits coverage, analysis and a manifest automatically.

A failing candidate returns to ascend-triton-operator-development. A fully passing matrix can proceed to ascend-triton-kernel-optimization.

Read the relevant detail only when needed:

- [behavior](references/behavior.md)
- [case design](references/case-design.md)

- [Business input example](references/inputs.md)