activation / torch-ext /torch_binding.cpp
iamwyldecat's picture
feat(rms-norm): Impl fused RMSNorm
f3b99fb
#include <torch/library.h>
#include "registration.h"
#include "torch_binding.h"
TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
// Activation ops
ops.def("poly_norm(Tensor! out, Tensor input, Tensor weight, Tensor bias, float eps) -> ()");
ops.def("poly_norm_backward(Tensor! input_grad, Tensor! weight_grad, Tensor! bias_grad, Tensor output_grad, Tensor input, Tensor weight, float eps) -> ()");
ops.impl("poly_norm", torch::kCUDA, &poly_norm);
ops.impl("poly_norm_backward", torch::kCUDA, &poly_norm_backward);
// Activation ops
ops.def("rms_norm(Tensor! out, Tensor input, Tensor weight, float eps) -> ()");
ops.def("rms_norm_backward(Tensor! input_grad, Tensor! weight_grad, Tensor output_grad, Tensor input, Tensor weight, float eps) -> ()");
ops.impl("rms_norm", torch::kCUDA, &rms_norm);
ops.impl("rms_norm_backward", torch::kCUDA, &rms_norm_backward);
}
REGISTER_EXTENSION(TORCH_EXTENSION_NAME)