Inside the tracer

Everything above describes tracing from the outside: what freezes, what disappears, what gets recompiled. Seeing the mechanism from the inside is worth the extra hour, because a description of a tracer is not the same thing as watching one run. The three resources below take you from prose to source to a corpus large enough to trust the pattern instead of one running example.

The autodidax notebook builds a working tracer from scratch, implementing the same machinery this chapter has been describing in prose: tracers standing in for arrays, constants getting frozen into constvars, a jaxprThe traced program: one equation per primitive in single-assignment form, every shape and dtype stated.taught in /l/jaxpr → accumulating equation by equation. Reading about that process only gets you so far. Watching it happen, one traced call at a time, is what makes tracing stop feeling unexplainable.

Once the mechanism is familiar, the jaxprThe traced program: one equation per primitive in single-assignment form, every shape and dtype stated.taught in /l/jaxpr → documentation is the reference worth keeping open. It defines the grammar this chapter has been using informally: constvars, invars, equations, the exact vocabulary a tracer emits. When a jaxpr dump looks unfamiliar later in the course, this page is what resolves it, not this chapter.

For scale, the corpus x-ray in the gym keeps seventeen real programs open at once, each with its traced output on screen. Skimming a handful of them is a fast way to check whether the rules from this chapter, frozen constants, vanished branches, shape-locked recompiles, hold up outside the one running example you've been tracing by hand. One running example proves the mechanism exists; seventeen prove it's a rule, not a coincidence of the example you happened to be given.