#include <torch/library.h>

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

TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def("fp4_quantize(Tensor x, Tensor(a!) resid, bool stochastic, int seed, int step) -> (Tensor, Tensor)");
  ops.def("fp4_dequantize(Tensor packed, Tensor scale, int M, int K) -> Tensor");

#if defined(CUDA_KERNEL) || defined(ROCM_KERNEL)
  ops.impl("fp4_quantize", torch::kCUDA, &fp4_quantize);
  ops.impl("fp4_dequantize", torch::kCUDA, &fp4_dequantize);
#endif
}

REGISTER_EXTENSION(TORCH_EXTENSION_NAME)