Activations stay on the device between layers (net/program.rs lowers the graph; every
operator is a WGSL kernel, one submission and read-back per run); faces of a photo go
in batches of up to 16; Settings > Faces shows each model's device and a CPU-only
switch (faces.gpu); the scan's pace limits GPU use; the GPU check is an activity row.
YuNet 58 -> 21 ms, SFace 40 -> 9.6 ms, AuraFace 295 -> 71 ms (heavily loaded machine).
Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An accelerator hook in the face interpreter takes Conv nodes; lightcraft-gpu implements it with a WGSL
implicit-GEMM convolution (any stride, padding, grouping; split-K for small maps; a direct kernel for depthwise
layers), each model's weights resident on a device of its own. After loading, detection and recognition check
the GPU in the background (agreement within 1e-3 and faster than the CPU) and otherwise stay on the CPU, which
remains the reference; the first GPU error drops it. Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
