Zero sized tensor support for repeat_interleave (#23717)

Summary:
Fixes pytorch/pytorch#22753
Pull Request resolved: pytorch/pytorch#23717

Differential Revision: D16623598

Pulled By: mrshenli

fbshipit-source-id: 297a3274fb5a5b2fcc0c3ad601337d7eb29fdca2

aten/src/ATen/native/Repeat.h

@@ -9,6 +9,9 @@ static inline Tensor repeat_interleave_common(const Tensor &repeats) {
     TORCH_CHECK(repeats.dim() == 1, "repeat_interleave only accept 1D vector as repeat");
     TORCH_CHECK(repeats.scalar_type() == at::kLong, "repeats has to be Long tensor");
     TORCH_CHECK((repeats >= 0).all().item<uint8_t>(), "repeats can not be negative");
+    if (repeats.size(0) == 0) {
+        return at::empty_like(repeats);
+    }
     Tensor repeats_ = repeats.contiguous();
     Tensor cumsum = repeats.cumsum(0);
     int64_t total = cumsum[-1].item<int64_t>();