Fix failing test due to a bug in NumPy when using OpenBLAS (#67679)

Summary:
Pull Request resolved: #67679

implementations

Fixes #67675

cc mruberry

Test Plan: Imported from OSS

Reviewed By: anjali411

Differential Revision: D32368698

Pulled By: mruberry

fbshipit-source-id: 3ea6ebc43c061af2f376cdf5da06884859bbbf53

test/test_linalg.py

@@ -1638,7 +1638,12 @@ def run_test_case(input, p):
         a = torch.eye(3, dtype=dtype, device=device)
         a[-1, -1] = 0  # make 'a' singular
         for p in norm_types:
-            run_test_case(a, p)
+            try:
+                run_test_case(a, p)
+            except np.linalg.LinAlgError:
+                # Numpy may fail to converge for some BLAS backends (although this is very rare)
+                # See the discussion in https://github.com/pytorch/pytorch/issues/67675
+                pass
 
         # test for 0x0 matrices. NumPy doesn't work for such input, we return 0
         input_sizes = [(0, 0), (2, 5, 0, 0)]