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

Hugging Face Blog · 1 month ago

PyTorch's profiler traces show that nn.Linear uses an optimized cuBLAS kernel (addmm) that folds bias addition into the matrix multiplication via an epilogue, eliminating the need for a separate add kernel. When torch.compile optimizes a single Linear layer, it removes only CPU dispatch overhead by pre-computing strides at compile time—the GPU kernel remains identical. In a three-layer MLP with GeGLU activation, the forward pass launches exactly five GPU kernels: three GEMMs for the linear projections plus one each for GeLU and element-wise multiplication.