What problem does it solve? Rewriting PyTorch operators to run on Ascend NPU hardware requires writing Triton kernels, preserving the original API contract, and proving numerical correctness — a process that is error-prone when done ad hoc. This Skill provides a disciplined conversion workflow with mandatory validation gates. ## Core Features & Use Cases - PyTorch-to-Triton Conversion: Transforms a PyTorch operator file into a Triton Ascend NPU kernel-backed operator while keeping the public PyTorch-facing API intact. - Enforced Validation Gate: Requires differential or standalone validation through the triton-npu-optimize skill's run-eval commands, forbidding ad-hoc comparison scripts or self-declared passes. - Strict Kernel Purity Rules: Ensures all computation lives inside @triton.jit kernels, with forward() limited to buffer allocation and kernel launches. - Use Case: Given a PyTorch operator implementing a fused reduction, convert it to a Triton NPU kernel file, preserve its input-helper block, and validate it differentially against the original implementation. ## Quick Start Convert my PyTorch operator file into a Triton NPU-backed operator and validate it with differential testing against the original.
