Tutorials

Seven notebooks, each building on the last: write and run a kernel, pick the right plane type, write a kernel that stops itself, take gradients backward and forward, read/scatter/reduce across a batch, and differentiate through branches and stops.

• Run your first kernel
• Arguments and shapes
• Stop when you’re done
• Gradients, backward
• Gradients, forward (and what autodiff refuses)
• Lookups, scatters and sums
• Gradients through branches and stops

Next: Vocabulary — declare a kernel’s own vocabulary once and reuse it across a whole family of models.