From Operator to Model: Registration, Porting, and Backend Integration

2.20.1PyTorch Custom Operator, Autograd, and Dispatcher#

2.20.2Choosing Among Operator Library Calls, Compiler Fusion, and Custom Kernels#

2.20.3Matching Shape, Layout, Precision, Sparse Format, and Device Capability#

2.20.4JIT Cache, Precompiled Kernels, and Architecture-Specific Artifacts#

2.20.5Graph Capture, Asynchronous Execution, and Custom Operator Compatibility#

2.20.6Distinguishing API Compatibility, Numerical Equivalence, and Performance Portability#

2.20.7End-to-End Integration from GEMM to Attention to Transformer Block#