PyTorch 2.6.0 Compatibility Fix for Kaolin

[ansorre]

Issue Summary

Kaolin currently encounters compilation errors with PyTorch 2.6.0 due to API changes in PyTorch. However, with minimal modifications to just three files, Kaolin can be successfully built and used with PyTorch 2.6.0.

Error Details

When attempting to build Kaolin with PyTorch 2.6.0, several CUDA compilation errors occur with the message:

error: no suitable conversion function from "const at::DeprecatedTypeProperties" to "c10::ScalarType" exists

This is due to the deprecation of the behavior of .type() in PyTorch 2.6.0, which previously could be used with AT_DISPATCH_FLOATING_TYPES_AND_HALF macro.

Solution

The fix is straightforward and involves modifying just three files by replacing all occurrences of .type() with .scalar_type():

1. kaolin/csrc/ops/spc/query_cuda.cu
2. kaolin/csrc/ops/spc/point_utils_cuda.cu
3. kaolin/csrc/render/spc/raytrace_cuda.cu

For example, in query_cuda.cu, lines like:

AT_DISPATCH_FLOATING_TYPES_AND_HALF(query_coords.type(), "query_cuda", ([&] {
    // Implementation
}));

Should be changed to:

AT_DISPATCH_FLOATING_TYPES_AND_HALF(query_coords.scalar_type(), "query_cuda", ([&] {
    // Implementation
}));

Testing

After applying these changes, Kaolin successfully compiles and installs with PyTorch 2.6.0 on Windows, and all Kaolin nodes in ComfyUI load without errors.

Recommendation

It would be beneficial to update the Kaolin codebase to use .scalar_type() instead of .type() for future compatibility with PyTorch, as this method is the recommended approach moving forward.