FAQ

Why is the kernelization step needed?

In earlier versions of kernels, a layer’s forward method was replaced by use_kernel_forward_from_hub and replace_kernel_forward_from_hub. The new forward would dispatch to a kernel based on the device type, whether a model was training, etc. However, this approach was fundamentally incompatible with torch.compile since it relied on data-dependent branching.

To avoid branching, we have to make dispatch decisions ahead of time, which is what the kernelize function does.

< > Update on GitHub