adding error message when attempting to use Large tensor with linalg_syevd

src/operator/tensor/la_op.h

@@ -475,6 +475,8 @@ inline bool LaEigFactShape(const nnvm::NodeAttrs& attrs,
   const mxnet::TShape& in_a = (*in_attrs)[0];
   const mxnet::TShape& out_u = (*out_attrs)[0];
   const mxnet::TShape& out_l = (*out_attrs)[1];
+  CHECK_LE(in_a.Size(), INT_MAX)
+    << "Large tensors are not supported by Linear Algebra operator syevd";
   if ( in_a.ndim() >= 2 ) {
     // Forward shape inference.
     const int ndim(in_a.ndim());

tests/nightly/test_large_array.py

@@ -27,7 +27,8 @@
 
 from mxnet.test_utils import rand_ndarray, assert_almost_equal, rand_coord_2d, default_context, check_symbolic_forward, create_2d_tensor, get_identity_mat, get_identity_mat_batch
 from mxnet import gluon, nd
-from common import with_seed, with_post_test_cleanup
+from common import with_seed, with_post_test_cleanup, assertRaises
+from mxnet.base import MXNetError
 from nose.tools import with_setup
 import unittest
 
@@ -1352,6 +1353,16 @@ def run_trsm(inp):
     check_batch_trsm()
 
 
+def test_linalg_errors():
+    def check_syevd_error():
+        A = get_identity_mat(LARGE_SQ_X)
+        for i in range(LARGE_SQ_X):
+            A[i,i] = 1
+        assertRaises(MXNetError, mx.nd.linalg.syevd, A)
+
+    check_syevd_error()
+
+
 def test_basic():
     def check_elementwise():
         a = nd.ones(shape=(LARGE_X, SMALL_Y))