#include <torch/library.h>

#include "registration.h"
#include "torch_binding.h"


TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
    ops.def("residual_rms(Tensor input, Tensor residual, Tensor weight, Tensor scale_tensor, float epsilon, Tensor! output, Tensor next_buffer, int num_threads, bool force_scalar) -> ()");
    ops.impl("residual_rms", torch::kCUDA, &residual_rms);
}

REGISTER_EXTENSION(TORCH_EXTENSION_NAME)