diff --git a/src/zarr/core/indexing.py b/src/zarr/core/indexing.py
index 3d47f5f183..ca227be094 100644
--- a/src/zarr/core/indexing.py
+++ b/src/zarr/core/indexing.py
@@ -1346,8 +1346,15 @@ def decode_morton(z: int, chunk_shape: ChunkCoords) -> ChunkCoords:
 
 
 def morton_order_iter(chunk_shape: ChunkCoords) -> Iterator[ChunkCoords]:
-    for i in range(product(chunk_shape)):
-        yield decode_morton(i, chunk_shape)
+    i = 0
+    order: list[ChunkCoords] = []
+    while len(order) < product(chunk_shape):
+        m = decode_morton(i, chunk_shape)
+        if m not in order and all(x < y for x, y in zip(m, chunk_shape, strict=False)):
+            order.append(m)
+        i += 1
+    for j in range(product(chunk_shape)):
+        yield order[j]
 
 
 def c_order_iter(chunks_per_shard: ChunkCoords) -> Iterator[ChunkCoords]:
diff --git a/tests/test_codecs/test_codecs.py b/tests/test_codecs/test_codecs.py
index dfb8e1c595..2025e72937 100644
--- a/tests/test_codecs/test_codecs.py
+++ b/tests/test_codecs/test_codecs.py
@@ -204,6 +204,26 @@ def test_morton() -> None:
     ]
 
 
+@pytest.mark.parametrize(
+    "shape",
+    [
+        [2, 2, 2],
+        [5, 2],
+        [2, 5],
+        [2, 9, 2],
+        [3, 2, 12],
+        [2, 5, 1],
+        [4, 3, 6, 2, 7],
+        [3, 2, 1, 6, 4, 5, 2],
+    ],
+)
+def test_morton2(shape) -> None:
+    order = list(morton_order_iter(shape))
+    for i, x in enumerate(order):
+        assert x not in order[:i]  # no duplicates
+        assert all(x[j] < shape[j] for j in range(len(shape)))  # all indices are within bounds
+
+
 @pytest.mark.parametrize("store", ["local", "memory"], indirect=["store"])
 def test_write_partial_chunks(store: Store) -> None:
     data = np.arange(0, 256, dtype="uint16").reshape((16, 16))
diff --git a/tests/test_codecs/test_sharding.py b/tests/test_codecs/test_sharding.py
index 78f32fef0e..51c82067f3 100644
--- a/tests/test_codecs/test_sharding.py
+++ b/tests/test_codecs/test_sharding.py
@@ -393,3 +393,32 @@ async def test_sharding_with_empty_inner_chunk(
     print("read data")
     data_read = await a.getitem(...)
     assert np.array_equal(data_read, data)
+
+
+@pytest.mark.parametrize("store", ["local", "memory"], indirect=["store"])
+@pytest.mark.parametrize(
+    "index_location",
+    [ShardingCodecIndexLocation.start, ShardingCodecIndexLocation.end],
+)
+@pytest.mark.parametrize("chunks_per_shard", [(5, 2), (2, 5), (5, 5)])
+async def test_sharding_with_chunks_per_shard(
+    store: Store, index_location: ShardingCodecIndexLocation, chunks_per_shard: tuple[int]
+) -> None:
+    chunk_shape = (2, 1)
+    shape = [x * y for x, y in zip(chunks_per_shard, chunk_shape, strict=False)]
+    data = np.ones(np.prod(shape), dtype="int32").reshape(shape)
+    fill_value = 42
+
+    path = f"test_sharding_with_chunks_per_shard_{index_location}"
+    spath = StorePath(store, path)
+    a = Array.create(
+        spath,
+        shape=shape,
+        chunk_shape=shape,
+        dtype="int32",
+        fill_value=fill_value,
+        codecs=[ShardingCodec(chunk_shape=chunk_shape, index_location=index_location)],
+    )
+    a[...] = data
+    data_read = a[...]
+    assert np.array_equal(data_read, data)