3 ms·
The gather-scatter sparse conv should be fairly generic. Any model using 3x3x3 or 5x5x5 sparse convolutions on voxel grids could use it directly. The main thin
by shivampkumar 5mo ago
The gather-scatter sparse conv should be fairly generic. Any model using 3x3x3 or 5x5x5 sparse convolutions on voxel grids could use it directly.
The main thing that's TRELLIS-specific is the neighbor cache key format, but that's a few lines to adapt.
The SDPA attention swap is even more reusable - it's just padding variable-length sequences into batches and calling torch.nn.functional.scaled_dot_product_attention.