This series aims to explain everything a PyTorch operation does, from Python to the chip. Part 0 provides a map of eight levels, called floors, between your keyboard and the hardware.

The first surprise: the Python you write is the smallest layer. `torch.randn` is a compiled function from a 49 KB stub. The real PyTorch body is 235 MB of compiled shared libraries (`libtorch_cpu.dylib`, `libtorch_python.dylib`). Calling `import torch` loads that body.

Crossing from Python to C++ costs time. The smallest operation, a one-element add, takes 0.54 microseconds, almost pure crossing cost. Deeper inside, the dispatcher routes every operation through a fixed stack of layers (e.g., autograd, mixed precision) before choosing a concrete kernel for the device and data type.

This build has 3,677 registered operation names, each mapping to many kernels. The dispatcher's final job is to pick exactly one cell per call. The rest of the series will explain each layer in detail.