diff --git a/torch_xla/csrc/tensor_impl.h b/torch_xla/csrc/tensor_impl.h index 2e76fe6e8289..8a0fa9862a6d 100644 --- a/torch_xla/csrc/tensor_impl.h +++ b/torch_xla/csrc/tensor_impl.h @@ -8,6 +8,15 @@ namespace torch_xla { +// [Note: Re-using upstream TensorImpl] As part of the LTC migration effort, +// we tried to re-use the upstream LTCTensorImpl +// (torch/csrc/lazy/core/tensor_impl.h) instead of having our own version of +// XLATensorImpl. However, LTCTensorImpl defines its own set of hard-coded +// dispatch keys in its constructor and has functions that call this +// constructor. In addition, updating XLATensorImpl to extend LTCTensorImpl does +// not produce much benefit since that wouldn't remove much duplicate code. As a +// result, we decided to keep and use XLATensorImpl. + // Tensor implementation class used to be fed to the at::Tensor. // Its scope is just to handle an XLATensor. class XLATensorImpl : public c10::TensorImpl {