ascend-triton-operator-development · git:20260911.9f92063 · 2026-09-11 · sha256 46ea4b4ef8eae17a
ascend-triton-operator-development git:20260911.9f92063A
Immutable. This exact content is served forever at /api/v1/blob/46ea4b4ef8eae17a.
--- name: ascend-triton-operator-development description: Develop a first correct Ascend Triton operator from a PyTorch reference or migrate an existing GPU Triton kernel to Ascend, including semantic audit, explicit task contracts, hardware-aware grid and tiling design, implementation, and handoff to correctness validation. Use for new kernel implementation, CUDA/GPU Triton migration, or repairing a candidate that has not yet passed correctness. Do not use for a kernel that already passes all planned cases and only needs performance tuning, for isolated torch_npu or ACLNN debugging, or for model-level graph failures. --- # ascend-triton-operator-development Implement a first correct Ascend Triton operator or migrate an existing GPU Triton kernel. Resolve semantics from the reference and callers before selecting a grid or tile. Separate logical shape from physical layout and reductions. GPU launch assumptions need an Ascend-specific design; choose a simple correct candidate before tuning. ## 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-operator-development/scripts/triton_development.py --config operator.json --kernel kernel.py --validation-manifest validation/manifest.json ``` The business config contains op_name, mode, source, reference, target, cases and tolerances. The report consumes the actual kernel and validation manifest, checking kernel identity and passing case coverage. Optional --semantic-report and --sketch attach useful design artifacts. Run ascend-triton-kernel-validation for the candidate. Continue to optimization only after the planned correctness cases pass. Read the relevant detail only when needed: - [behavior](references/behavior.md) - [semantic review](references/semantic-review.md) - [architecture and codegen](references/architecture-and-codegen.md) - [Business input example](references/inputs.md)