pattern-dev

What problem does it solve? Writing graph fusion patterns for PyTorch Inductor's pattern matcher often fails silently: patterns register and pass unit tests but never match the real compiled model graph, or they match yet deliver no wall-clock benefit. This Skill provides a disciplined Phase 1-7 lifecycle for creating, registering, debugging, and validating fusion patterns on torch + torch_npu NPU environments. ## Core Features & Use Cases - Pattern lifecycle management: Covers model analysis, pattern creation via register_replacement or GraphPatternEntry manual graph rewriting, three-file registration, unit verification, mismatch debugging, integration validation, and Copy kernel elimination. - Mismatch diagnosis: A catalog of 7 real mismatch types (decomposition differences, placeholder vs get_attr, scalar literal drift) plus graph dump comparison methods and real-graph probe scripts. - Benefit root-cause analysis: Diagnoses hit-but-no-benefit cases (dtype promotion, input materialization, format conversion, fallback mismatches, inefficient custom kernels) with kernel diff tooling. - Use Case: When a user says "my compile fusion is not hitting" or "InplaceCopy kernels doubled after compile", this Skill walks them through dumping the real FX graph, aligning nodes against the pattern, and confirming fusion via kernel_details.csv counts. ## Quick Start Ask the assistant to help write or debug an Inductor fusion pattern for a model, for example to create an RMSNorm fusion pattern and verify it actually hits in the compiled graph.