What problem does it solve? Integrating a verified SHMEM operator into PyTorch requires writing C++ CustomClassHolder bindings, CMake build configuration, and multi-PE test scripts by hand, which is error-prone around dtype dispatch, ABI flags, and shared-memory session management. ## Core Features & Use Cases - C++ Binding Generation: Produces torch_bind_<op_name>.cpp with a CustomClassHolder subclass, TORCH_CHECK validation, dtype dispatch, and REGISTER_SHMEM_OPS_CLASS registration, plus the shared Manager class when missing. - CMake Integration: Adds the operator to the shared custom-ops/torch_binding/ build producing shmem_custom_ops_torch.so, or to the in-tree examples/torch_binding/ producing aclshmem_torch.so, with correct _GLIBCXX_USE_CXX11_ABI detection. - Multi-PE Test Generation: Creates torch_test_<op_name>.py with fixed-seed data generation, golden output comparison, and multiprocessing-based 2-PE and 8-PE verification on Ascend NPU. - Use Case: After a SHMEM AllGather-style operator passes correctness evaluation, use this Skill to generate its Torch binding, compile the shared extension, and run an 8-PE PyTorch test to confirm end-to-end correctness. ## Quick Start Generate the PyTorch CustomClass binding and multi-PE test script for my verified SHMEM operator, then compile the shared Torch extension and run the 8-PE validation.
