commit 8b8c76e6f381d572c8da9b5e10ab677376ee357b
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jun 17 10:08:32 2026 +0200

dev-python/spconv-cu126: new package, add 2.3.8

Sparse-convolution backend for the TRELLIS image-to-3D node. The TRELLIS
pretrained weights use spconv's weight naming and layout, so spconv (not
torchsparse) is the backend that can load them. The cu126 binary wheel bundles
its own CUDA kernels via cumm and does not link libtorch, so it runs against
this host's CUDA 13.3 / torch 2.11 stack. cumm is capped <0.8.0 by upstream.