#include <torch/library.h>

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

// Registers `flash_attention_in_cuda_from_scratch` as a custom op so it is callable from Python as
// torch.ops.flash_attention_in_cuda_from_scratch.flash_attention_in_cuda_from_scratch(...). Adjust the schema to your kernel.
TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  ops.def("flash_attention_in_cuda_from_scratch(Tensor! out, Tensor input) -> ()");
  ops.impl("flash_attention_in_cuda_from_scratch", torch::kCUDA, &flash_attention_in_cuda_from_scratch);
}

REGISTER_EXTENSION(TORCH_EXTENSION_NAME)