Torch.empty - how is it implemented

I am starting to getting more accustomed with pytorch source code but I’ve recently hit a wall. I don’t quite understand C <-> CUDA <-> Python bindings yet I guess and that’s my problem here.

I just couldn’t find the source code of torch.empty function. Probably the easiest way to find it out is to ask :wink:


1 Like

For PyTorch 1.17 or 1.18, do you know how the ‘empty(…)’ implemented?

I just found :

Tensor empty(
    IntArrayRef size,
    at::optional<DimnameList> names,
    const TensorOptions& options,
    optional<MemoryFormat> optional_memory_format) {
  if (!names.has_value()) {
    return at::empty(size, options, optional_memory_format);
  TORCH_CHECK(options.layout() == Layout::Strided,
      "NYI: named tensors only support strided layout");
  TORCH_CHECK(options.device().is_cpu() || options.device().is_cuda(),
      "NYI: named tensors only support CPU and CUDA tensors");
  auto result = at::empty(size, options, optional_memory_format);
  internal_set_names_inplace(result, names);
  return result;

same here. Wasn’t able to find the implementation for at::empty