diff --git a/autoray/autoray.py b/autoray/autoray.py
index 386177e..82e5064 100644
--- a/autoray/autoray.py
+++ b/autoray/autoray.py
@@ -4124,7 +4124,8 @@ def torch_pad(array, pad_width, mode="constant", constant_values=0):
     try:
         # numpy takes pads like ((0, 0), (1, 1), ... (n-1, n-1))
         # torch takes pads like (n-1, n-1, n-2, n-2, n-3, n-3, ...)
-        pad = tuple(itertools.chain.from_iterable(pad_width))[::-1]
+        # i.e. axes reversed, but each (before, after) pair kept in order
+        pad = tuple(itertools.chain.from_iterable(reversed(pad_width)))
 
         # a single tuple was specified ((a, b),) - use for all axes
         if len(pad) == 2:
diff --git a/tests/test_autoray.py b/tests/test_autoray.py
index 6614fe3..b8f88d7 100644
--- a/tests/test_autoray.py
+++ b/tests/test_autoray.py
@@ -418,12 +418,18 @@ def test_pad(backend):
         (((1, 2),), (6, 7, 8)),
         # different pad for every axis
         (((4, 3), (2, 4), (3, 2)), (10, 10, 10)),
+        # asymmetric, including zero pads
+        (((0, 1), (2, 0), (3, 1)), (4, 6, 9)),
     ]:
         B = ar.do("pad", A, pad_width)
         assert shape(B) == new_shape
         assert ar.to_numpy(ar.do("sum", A)) == pytest.approx(
             ar.to_numpy(ar.do("sum", B))
         )
+        # check the padding is placed the same way numpy places it
+        np.testing.assert_allclose(
+            ar.to_numpy(B), np.pad(ar.to_numpy(A), pad_width)
+        )
 
 
 @pytest.mark.parametrize("backend", gen_params(backends=...))