the whole stack · one map · four courses across it
One program, all the way down.
Two frameworks converge on one portable graph, cross the runtime seam under it, and come out as code for two different machines. Every chapter on this site is a floor or a seam of that descent. Walk far enough down and unfamiliar dumps stop being unfamiliar, which is the point: the courses end with kernels you derived, tested, and measured against the compiler.
no floor finished yet
Depth counts floors, not chapters. A floor lands when every unit you opened on it is finished, and 4/5 gates have passed on real hardware so far.
03/06step 2 · one pass over K1, V1: rescale (m, l, acc), never materializing scores to HBM
the map · each cell holds the units that live on that floor and the equipment that serves them · pick a course to light its route
- PyTorch side not yet
- JAX side not yet
- TPU side not yet
- L0 the program you wrote
- L1 the traceFX / lazy IR
- L2 the portable graph
- the seam · PJRT · the program changes owners here
- L3 the compiler
- L4 the kernelTriton / GPU nothing here yetthe gap
- L5 machine codePTX / SASS nothing here yet
- L6 the siliconSM / HBM nothing here yet
- L7 the podNVLink nothing here yet
The workshop
Drills, reproduced failures and measured numbers all sit on the map now, under the floor they serve. These six instruments are the ones to open first, and ⌘K jumps to anything else.
Gates pass only on real hardware, and failures stay published. Every value on this site states what it is: a roofline prediction, an off-chip verification, or a measured run with chip, dtype, shapes, and date attached.