. How a Model Becomes GPU Code
The path from a traced PyTorch model to CUDA kernels, and where the compiler has to choose between correct answers.
. Passes and Rewrite Rules
How one compiler rewrite happens — pattern, match, replacement — and what the pipeline is made of.
. Forks and Knobs
How a compiler choice is represented before anything is built — knobs, fork trees, schedule keys, and pinning.
. Measuring and Recalling
The two ways a fork is answered — measure it now, or recall what was measured before — and the four places knowledge is kept.
. Inside a Tuning Run
What emmy tune actually does — a search over forks, split into two levels, with every kernel tuned on its own.
. The Deploy Evidence Hierarchy
How a compile answers a fork — measured evidence first, then prediction, then the rule's first option.
. Golden Configurations
The reviewed, per-GPU measurements that ship with the repository — how they are recorded, and how a recording replays exactly.
. Inside the Prior
The model that ranks schedules before they are measured — its features, its two halves, how it is trained, and the check that decides whether it is trusted.
. Storage, Checks and Limits
How measurements are keyed and stored, how the prior is evaluated when it chooses badly, and where the whole design falls short.