| TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) { | |
| ops.def("muon_orthogonalize(Tensor Xb, int steps, float eps) -> Tensor"); | |
| ops.def("muon_sr_update(Tensor(a!) param, Tensor O, float lr, float scale, float wd, " | |
| "int seed, int step, int base_index) -> ()"); | |
| ops.impl("muon_orthogonalize", torch::kCUDA, &muon_orthogonalize); | |
| ops.impl("muon_sr_update", torch::kCUDA, &muon_sr_update); | |
| } | |
| REGISTER_EXTENSION(TORCH_EXTENSION_NAME) | |