All compute (WASM + WebGPU) runs in a dedicated Web Worker — page stays
fully responsive during the entire inference. Whole-body CT at 3 mm.
·
← v1 (main-thread)
✓ 100.0000% byte-exact canonical-argmax conformance with
nnUNetv2_predict CPU (task 297). Uses symmetric
pad_nd_image + GPU-side sliding_unpad_accum
kernel so postprocess_gpu's per-channel logit resample
sees the correct (unpadded) canonical shape.