jax floor
serves jax-path chapters 02 · 05 · streaks check themselves off; the same skill counts on every floor that drills it · all floors
Read the recording until it is boring.
The jax path's claim is that a jaxpr is readable on sight, and reading is a drilled skill. These stations run on the same generated corpus the kernel floor uses: seventeen real programs, traced and dumped, answers computed from the structured facts, never invented. Streaks persist in your browser and nowhere else.
Active recall, one tap
- bfloat16[64, 128]
- bfloat16[128, 512]
what shape and dtype comes out?
{ lambda ; a:bf16[32,64] b:bf16[64,64] c:bf16[64]. let d:bf16[32,64] = dot_general[ dimension_numbers=(([1], [0]), ([], [])) preferred_element_type=bfloat16 ] a b e:bf16[1,64] = broadcast_in_dim[ broadcast_dimensions=(1,) shape=(1, 64) sharding=None ] c f:bf16[32,64] = add d e g:bf16[32,64] = tanh f in (g,) }
x-rayed
Seventeen real programs, wired together
Pick a program and hover any column: the source line, the jaxpr it traced to, and the StableHLO it lowered to light up as one. This is jax chapter 02's mastery bar made physical: narrate a program you have not seen, then another. When the middle column stops surprising you, the chapter has done its work.
the transformer FFN: two matmuls around a GELU; watch bias broadcasts appear
bright lines are mapped equations · dim lines are scaffolding, compiler-made constants and broadcasts, or alignments the mapper skips rather than guesses
hover any line, bright or dim: it explains itself here · source lines recorded by jax itself, IR mapping computed mechanically