What problem does it solve? Manually identifying which parts of a PyTorch model can be replaced with torch_npu fused operators on Ascend NPU is error-prone: developers must decompose Attention, MoE, FFN, and Norm modules, match them against a large operator catalog, and verify API signatures, dtype, shape, layout, and cache constraints before replacing code. This Skill provides a structured analyze-match-replace-validate workflow grounded in local torch_npu docstring queries and repository reference implementations. ## Core Features & Use Cases - Module decomposition and matching: Breaks models into replaceable sub-chains (RoPE, KV Cache write, Flash Attention, MoE gating/routing/dispatch, Residual+Norm) and matches each against repository reference implementations for GQA, MLA Absorb, and MLA+Indexer paths. - API verification tooling: Ships a query script that reads torch_npu's built-in _op_plugin_docs.py to show, search, and list operator documentation, with fallback entries and an offline operator catalog. - Use Case: Given a GQA model whose Prefill uses current q/k/v and Decode reads from a PA KV Cache, the Skill guides separate analysis of each stage, verifies npu_fused_infer_attention_score constraints via local docstrings, then replaces and validates accuracy and performance module by module. ## Quick Start Analyze my PyTorch model code and tell me which modules can be replaced with torch_npu fused operators, with candidate APIs and a validation plan.
