FindDefinition

Firstly subm conv isn't equivalent to dense conv. All spconv weights use [out channel, H, W, input vhannel], torch conv may use [ out channel, input channel, H, W].

You need to keep point-to-voxel indices when do point2voxel, then you can use it to recover point cloud. You can use generate_voxel_with_id from PointToVoxel to get that tensor.

@lkhphuc there are two dispatch in pytorch context parallel impl: 1. dispatch native op (e.g. `aten._scaled_dot_product_flash_attention.default`) when input is `DTensor`, this dispatch only happens when sdpa native op is invoked...

their backward implementation is slower than simple warp-reduce backward in my codebase (not gsplat, backward algo is similar), don't need to waste time on that. three code segments that have...

look forward to more generic N-D parallel (device mesh, TP, CP) support instead of fsdp2 only. I have implemented a simple `AcceleratorNd` by inheritance to support this, but I found...