Compatibility notes

Torch Lattice compatibility is about three layers: original TorchSparse migration, Torch/CUDA runtime compatibility, and MLX artifact compatibility.

Original TorchSparse

The project is a fork lineage, not a drop-in promise. Covered migration behavior is tested through the migration CLI. Code that depends on implicit stride-1 convolution semantics should be rewritten to explicit SubmConv3d or Conv3d according to the intended support behavior.

Torch/CUDA

The package currently targets a modern PyTorch/CUDA stack. Keep PyTorch, CUDA runtime, NVCC, and driver versions aligned. Mismatched CUDA installations are the most common source of build failures.

MLX artifacts

Artifact compatibility is the deployment contract. If a model exports and replays through conformance with acceptable numerical error, it is compatible at the level that matters for MLX deployment.