Add meta-functions for asynchronous_*_cumsum ops (pytorch#2028)

Summary:

ATT

Differential Revision: D49467255

Author: aakhundov

fbgemm_gpu/src/sparse_ops/sparse_ops_meta.cpp

@@ -46,12 +46,26 @@ Tensor batched_unary_embeddings_forward_meta(
   return at::empty_symint({N, B, T}, weight.options());
 }
 
+Tensor asynchronous_inclusive_cumsum_meta(const Tensor& t_in) {
+  return at::empty_symint(t_in.sym_sizes(), t_in.options());
+}
+
+Tensor asynchronous_exclusive_cumsum_meta(const Tensor& t_in) {
+  return at::empty_symint(t_in.sym_sizes(), t_in.options());
+}
+
 } // namespace
 
 } // namespace fbgemm_gpu
 
 TORCH_LIBRARY_IMPL(fbgemm, Meta, m) {
   m.impl("pack_segments", TORCH_FN(fbgemm_gpu::pack_segments_forward_meta));
+  m.impl(
+      "asynchronous_inclusive_cumsum",
+      TORCH_FN(fbgemm_gpu::asynchronous_inclusive_cumsum_meta));
+  m.impl(
+      "asynchronous_exclusive_cumsum",
+      TORCH_FN(fbgemm_gpu::asynchronous_exclusive_cumsum_meta));
   m.impl(
       "asynchronous_complete_cumsum",
       TORCH_FN(fbgemm_gpu::asynchronous_complete_cumsum_meta));