What problem does it solve? When torch_fl must move to a different PyTorch minor version (e.g. creating a 2.9 branch or bumping to 2.11/2.12), the generated ATen bindings go stale and import torch_fl fails with "Mismatch in kernel C++ signatures". This Skill provides the hardware-independent workflow to pin a CPU-only torch, regenerate codegen against the new native_functions.yaml, and reconcile per-operator IListRef/ArrayRef signature rules. ## Core Features & Use Cases - Version branch management: Creates sibling per-torch-minor branches from main rather than inheriting stale signature fixes from other version branches. - ATen codegen regeneration: Runs scripts/codegen/codegen_ops.py against the newly pinned torch, with checks for dropped operators, failed FlagGems imports, and generator idempotency. - Signature reconciliation loop: Fixes the ArrayRef/IListRef rule in the generator (never hand-editing generated files) and iterates one operator at a time against the real import error. - Use Case: After bumping torch from 2.10 to 2.12, import torch_fl aborts on aten::cat with an IListRef/ArrayRef mismatch; use this Skill to re-derive the rule from the installed torchgen, regenerate, and verify the integration test suite. ## Quick Start Port torch_fl to PyTorch 2.9 by creating a new version branch, pinning a CPU-only torch 2.9, regenerating the ATen codegen, and fixing any kernel signature mismatches until import succeeds.
