torch-ext/torch_binding.cpp

#include "torch_binding.h"

#include <torch/library.h>

#include "registration.h"

TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  // poly_norm
  ops.def("poly_norm(Tensor! out, Tensor input, Tensor weight, Tensor bias, "
          "float eps) -> ()");
  ops.impl("poly_norm", torch::kCUDA, &poly_norm);

  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_backward", torch::kCUDA, &poly_norm_backward);

  // rms_norm
  ops.def(
      "rms_norm(Tensor input, Tensor weight, float eps) -> (Tensor, Tensor)");
  ops.impl("rms_norm", torch::kCUDA, &rms_norm);

  ops.def("rms_norm_backward(Tensor output_grad, Tensor input, Tensor weight, "
          "Tensor inv_rms, float eps) -> (Tensor, Tensor)");
  ops.impl("rms_norm_backward", torch::kCUDA, &rms_norm_backward);

  // fused_mul_poly_norm
  ops.def("fused_mul_poly_norm(Tensor! out, Tensor input, Tensor mul, Tensor "
          "weight, Tensor bias, "
          "float eps) -> ()");
  ops.impl("fused_mul_poly_norm", torch::kCUDA, &fused_mul_poly_norm);

  ops.def("fused_mul_poly_norm_backward(Tensor! input_grad, Tensor! mul_grad, "
          "Tensor! weight_grad, Tensor! "
          "bias_grad, Tensor output_grad, Tensor input, Tensor mul, Tensor "
          "weight, Tensor "
          "bias, float eps) -> ()");
  ops.impl("fused_mul_poly_norm_backward", torch::kCUDA,
           &fused_mul_poly_norm_backward);

  // fused_add_rms_norm
  ops.def("fused_add_rms_norm(Tensor input, Tensor residual, Tensor "
          "weight, float eps) -> (Tensor, Tensor)");
  ops.impl("fused_add_rms_norm", torch::kCUDA, &fused_add_rms_norm);

  ops.def(
      "fused_add_rms_norm_backward(Tensor output_grad, Tensor add_output_grad,"
      "Tensor input, Tensor weight, float eps, bool need_input_grad) -> "
      "(Tensor, Tensor)");
  ops.impl("fused_add_rms_norm_backward", torch::kCUDA,
           &fused_add_rms_norm_backward);

  // grouped_poly_norm (without scores, hidden_clamp < 0 = disabled)
  ops.def(
      "grouped_poly_norm_forward("
      "Tensor input, Tensor mul, Tensor weight, "
      "Tensor bias, Tensor offsets, "
      "float eps, int expert_offset, float hidden_clamp) -> (Tensor, Tensor)");
  ops.impl("grouped_poly_norm_forward", torch::kCUDA,
           &grouped_poly_norm_forward);

  ops.def("grouped_poly_norm_backward("
          "Tensor grad_output, Tensor input, Tensor mul, Tensor weight, "
          "Tensor bias, Tensor offsets, Tensor inv_rms, "
          "float eps, int expert_offset, float hidden_clamp) -> (Tensor, "
          "Tensor, Tensor, Tensor)");
  ops.impl("grouped_poly_norm_backward", torch::kCUDA,
           &grouped_poly_norm_backward);

  // grouped_poly_norm (with scores)
  ops.def(
      "grouped_poly_norm_forward_scored("
      "Tensor input, Tensor mul, Tensor weight, "
      "Tensor bias, Tensor offsets, Tensor scores, "
      "float eps, int expert_offset, float hidden_clamp) -> (Tensor, Tensor)");
  ops.impl("grouped_poly_norm_forward_scored", torch::kCUDA,
           &grouped_poly_norm_forward_scored);

  ops.def("grouped_poly_norm_backward_scored("
          "Tensor grad_output, Tensor input, Tensor mul, Tensor weight, "
          "Tensor bias, Tensor offsets, Tensor inv_rms, Tensor scores, "
          "float eps, int expert_offset, float hidden_clamp) -> (Tensor, "
          "Tensor, Tensor, Tensor, Tensor)");
  ops.impl("grouped_poly_norm_backward_scored", torch::kCUDA,
           &grouped_poly_norm_backward_scored);
}

REGISTER_EXTENSION(TORCH_EXTENSION_NAME)