A static-shape tensor IR, reverse-mode autodiff as a graph-to-graph transform,
a float64 interpreter that defines what every program means, and, from milestone 2, lowering
to WGSL compute kernels so a small model can train in the browser. No ML framework in the
engine; PyTorch is used offline only, to write golden files.
The question: when does kernel fusion pay off on an integrated
GPU, and when does it lose? That needs measured numbers, including the cases where fusion
is neutral or slower. Fused kernels are only worth timing once they are known to compute
the same thing as the unfused ones, so the first milestone is the reference everything
will be checked against. This page reports milestones 1 and 2 of 4, the hardening pass
between them (1.5), and the first half of milestone 3 (3a: training an MLP). The
transformer (3b) and the fusion study (4) are not done. Nothing earlier has been removed:
each milestone's record, failures and false alarms included, is further down, unchanged.