diff --git a/imperative/python/test/unit/core/test_indexing_op.py b/imperative/python/test/unit/core/test_indexing_op.py
index 6f483eb6934d2573bc20511cb197a7ffa35f5c79..7a77d239dfcfdc449ac220ba615ad733cfc61101 100644
--- a/imperative/python/test/unit/core/test_indexing_op.py
+++ b/imperative/python/test/unit/core/test_indexing_op.py
@@ -5,7 +5,6 @@ from tempfile import NamedTemporaryFile
 
 import numpy as np
 import pytest
-import torch
 from utils import make_tensor
 
 import megengine
@@ -624,25 +623,6 @@ def test_advance_indexing_with_bool(test_varnode):
     )
 
 
-def test_advance_indexing_autodiff():
-    x = Tensor([2, 2, 3, 4, 5, 6, 7, 8, 2], dtype="float32")
-    gm = GradManager()
-    gm.attach(x)
-    with gm:
-        a = x + 1
-        a[x > 3] = 0.3
-        b = a + 1
-        gm.backward(b.sum())
-    torch_x = torch.tensor(
-        [2, 2, 3, 4, 5, 6, 7, 8, 2], dtype=torch.float32, requires_grad=True
-    )
-    a = torch_x + 1
-    a[torch_x > 3] = 0.3
-    b = a + 1
-    (b.sum()).backward()
-    np.testing.assert_equal(x.grad.numpy(), torch_x.grad.numpy())
-
-
 @pytest.mark.parametrize("symbolic", [True, False, None])
 def test_subtensor_on_empty_tensor(symbolic):
     np_x = np.array([], dtype=np.float32).reshape(10, 0, 10)
diff --git a/imperative/python/test/unit/functional/test_tensor.py b/imperative/python/test/unit/functional/test_tensor.py
index 9d82fdafada715cc9c97d85d45733bb83d8c0b0f..42af686ee5b167712149f9cd55665ecbe3377ef8 100644
--- a/imperative/python/test/unit/functional/test_tensor.py
+++ b/imperative/python/test/unit/functional/test_tensor.py
@@ -227,80 +227,6 @@ def test_split_basic(is_varnode):
         set_symbolic_shape(saved_symbolic_shape)
 
 
-def test_concat_and_stack():
-    import copy
-    from megengine.autodiff import GradManager
-    import torch
-
-    def generate_test_data(max_nr_inp, max_dim, max_dim_len, test_concat=True):
-        nr_inp = np.random.randint(1, max_nr_inp) if max_nr_inp > 1 else 1
-        dims = np.random.randint(1, max_dim)
-        cat_axis = (
-            np.random.randint(-dims, dims)
-            if test_concat
-            else np.random.randint(-dims - 1, dims + 1)
-        )
-
-        ishape = [np.random.randint(0, max_dim_len) for _ in range(dims)]
-        ishapes = [copy.deepcopy(ishape) for _ in range(nr_inp)]
-        if test_concat:
-            for i in range(nr_inp):
-                ishapes[i][cat_axis] = np.random.randint(0, max_dim_len)
-
-        inp_nps = []
-        for ishape in ishapes:
-            inp_nps.append(np.random.randn(*ishape))
-        return inp_nps, cat_axis
-
-    def test_impl(max_nr_inp, max_dim, max_dim_len, test_concat):
-        inp_nps, cat_axis = generate_test_data(
-            max_nr_inp, max_dim, max_dim_len, test_concat
-        )
-        inp_mges = [Tensor(inp_np) for inp_np in inp_nps]
-        inp_torchs = [torch.tensor(inp_np, requires_grad=True) for inp_np in inp_nps]
-        if test_concat:
-            np_func, mge_func, torch_func = np.concatenate, F.concat, torch.cat
-        else:
-            np_func, mge_func, torch_func = np.stack, F.stack, torch.stack
-
-        res_np = np_func(inp_nps, axis=cat_axis)
-        grad_np = np.random.randn(*res_np.shape).astype(np.float32)
-
-        gm = GradManager().attach(inp_mges)
-        with gm:
-            res_mge = mge_func(inp_mges, axis=cat_axis)
-            gm.backward(res_mge, Tensor(grad_np))
-
-        res_torch = torch_func(inp_torchs, dim=cat_axis)
-        res_torch.backward(torch.tensor(grad_np))
-
-        np.testing.assert_allclose(res_mge.numpy(), res_torch.detach().cpu().numpy())
-        for inp_mge, inp_torch in zip(inp_mges, inp_torchs):
-            np.testing.assert_allclose(
-                inp_mge.grad.numpy(), inp_torch.grad.detach().cpu().numpy()
-            )
-
-    def test_concat(max_nr_inp, max_dim, max_dim_len):
-        test_impl(max_nr_inp, max_dim, max_dim_len, test_concat=True)
-
-    def test_stack(max_nr_inp, max_dim, max_dim_len):
-        test_impl(max_nr_inp, max_dim, max_dim_len, test_concat=False)
-
-    # test only one input
-    test_concat(1, 7, 16)
-    test_stack(1, 7, 16)
-
-    # test zero shape
-    test_concat(10, 7, 1)
-    test_stack(10, 7, 1)
-
-    for _ in range(3):
-        test_concat(10, 7, 16)
-
-    for _ in range(3):
-        test_stack(10, 7, 16)
-
-
 @pytest.mark.parametrize("symbolic", [None, False, True])
 def test_split(symbolic):
     x = Tensor(np.random.random((10, 20)), dtype=np.float32)