Disable SparseFillEmptyRows[Grad] on GPU
It breaks an internal workload with the error message "segment ids are not
increasing", which probably means that the output indices are not sorted in some
cases.
PiperOrigin-RevId: 371980850
Change-Id: Ie7eb322de5917fe55ee6318d262236e8db8da26d
diff --git a/tensorflow/core/kernels/sparse_fill_empty_rows_op.cc b/tensorflow/core/kernels/sparse_fill_empty_rows_op.cc
index de97e63..a2b8b07 100644
--- a/tensorflow/core/kernels/sparse_fill_empty_rows_op.cc
+++ b/tensorflow/core/kernels/sparse_fill_empty_rows_op.cc
@@ -263,7 +263,7 @@
#undef REGISTER_KERNELS
-#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
+#if 0 && (GOOGLE_CUDA || TENSORFLOW_USE_ROCM)
// The GPU implementation is async because it requires waiting for a
// host->device memcpy before the output is allocated (similar to
@@ -411,7 +411,7 @@
TF_CALL_NUMBER_TYPES(REGISTER_CPU_KERNELS);
#undef REGISTER_CPU_KERNELS
-#if GOOGLE_CUDA || TENSORFLOW_USE_ROCM
+#if 0 && (GOOGLE_CUDA || TENSORFLOW_USE_ROCM)
// Forward declarations of the functor specializations for GPU.
namespace functor {