Skip to main content

Trace One Op End to End

When you call torch.add, Python is just the doorbell — the real work happens in C++.

Key Insight​

A Python call like torch.add(a, b) does no math itself. It travels through the dispatcher, which picks the right kernel for your tensor's device and dtype, and lands in ATen, the C++ library that does the actual arithmetic.

Why This Matters​

Once you can follow one operation from the Python call all the way down to its CPU kernel, the framework stops feeling like magic. You can then trace any op and answer questions the documentation never covers.