Follow it from model weights to kernels, memory and serving, with lessons from building C-Kernel-Engine.
I made this for readers who want to look underneath the model interface. Start with a recording that exposed a compiler bug, then follow the questions I use to investigate correctness and speed.
Three pages from the guide. Open a preview for a closer look.
Put weights, generated calls, kernels, the server and agent harness on one page so you know which layer you are testing.
Follow the Whisper recording that returned almost one word, from its symptom to a wrong buffer connection and a measured fix.
Use the printable worksheet to record inputs, a correctness check, timing, and the next experiment rather than guessing at performance.
That run led us to a wrong buffer connection in the generated Whisper encoder. Finding the first bad value was a better lesson than simply saying the model worked.
If you can run a model and are curious about the code underneath, this is a place to begin. Each page points to deeper reading when you want to keep going. The examples come from CPU work in CKE, while also showing where GPU execution and shared infrastructure fit.
What am I running? Which values are wrong? Where does the time go? The printable worksheet gives you room to record what you find.
I write about the maths, the kernels, and what happens when I try to make them work. Join me for the investigations, including the bits that go wrong.
Get the illustrated field guide by email, followed by engineering notes as I write them.