#include <torch/library.h>

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

TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def("dequantize(Tensor blocks, int ggml_type, int rows, int cols, ScalarType dtype) -> Tensor");
  ops.def("mul_mat_vec(Tensor blocks, Tensor x, int ggml_type, int out_features) -> Tensor");

  // The schema is the same for every backend; only the implementation differs.
#if defined(CUDA_KERNEL) || defined(ROCM_KERNEL)
  ops.impl("dequantize", torch::kCUDA, &dequantize);
  ops.impl("mul_mat_vec", torch::kCUDA, &mul_mat_vec);
#elif defined(METAL_KERNEL)
  ops.impl("dequantize", torch::kMPS, &dequantize);
  ops.impl("mul_mat_vec", torch::kMPS, &mul_mat_vec);
#endif
}

REGISTER_EXTENSION(TORCH_EXTENSION_NAME)