Profiling in PyTorch (Part 2): From nn.Linear to a Fused MLP

Combining GPU operations can save memory traffic even when the main matrix calculations remain unchanged. This PyTorch tutorial follows a neural-network block in which an activation calculation writes an intermediate result to GPU memory and a multiplication immediately reads it back. Compilation combines those steps into one GPU program, keeping the intermediate value in faster storage on the chip.

In the tutorial's example, the three matrix multiplications still use the same underlying GPU programs. The improvement mechanism is removing an intermediate memory round trip. Building on the earlier profiling guide, it shows why inspecting what compilation changed is more informative than assuming every operation became faster.