diff --git a/imperative/python/megengine/functional/nn.py b/imperative/python/megengine/functional/nn.py
index a8f9397a02bc4be9be5cd4a3e2b8e913edb0cfe6..9c8e7d1e82619d8e34bf5ed4365a195e93c3b5f6 100644
--- a/imperative/python/megengine/functional/nn.py
+++ b/imperative/python/megengine/functional/nn.py
@@ -1071,13 +1071,12 @@ def interpolate(
         import numpy as np
         from megengine import tensor
         import megengine.functional as F
-        from megengine.test import assertTensorClose
 
         x = tensor(np.arange(1, 5, dtype=np.float32).reshape(1, 1, 2, 2))
         out = F.interpolate(x, [4, 4], align_corners=False)
         print(out.numpy())
         out2 = F.interpolate(x, scale_factor=2.)
-        assertTensorClose(out.numpy(), out2.numpy())
+        np.testing.assert_allclose(out.numpy(), out2.numpy())
 
     Outputs:
 
diff --git a/imperative/python/megengine/test/__init__.py b/imperative/python/megengine/test/__init__.py
deleted file mode 100644
index 44ed54c22e810586429e5aed6b0ae41da066e629..0000000000000000000000000000000000000000
--- a/imperative/python/megengine/test/__init__.py
+++ /dev/null
@@ -1,67 +0,0 @@
-# -*- coding: utf-8 -*-
-# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
-#
-# Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-import numpy as np
-
-
-def assertTensorClose(
-    v0, v1, *, max_err: float = 1e-6, allow_special_values: bool = False, name=None
-):
-    """
-    :param allow_special_values: whether to allow :attr:`v0` and :attr:`v1` to contain inf and nan values.
-    :param max_err: relative error
-    """
-    __tracebackhide__ = True  # pylint: disable=unused-variable
-
-    assert (
-        v0.dtype == v1.dtype
-    ), "Two Tensor must have same dtype, but the inputs are {} and {}".format(
-        v0.dtype, v1.dtype
-    )
-    v0 = np.ascontiguousarray(v0, dtype=np.float32).copy()
-    v1 = np.ascontiguousarray(v1, dtype=np.float32).copy()
-    if allow_special_values:
-        # check nan and rm it
-        v0_nan_mask = np.isnan(v0)
-        if np.any(v0_nan_mask):
-            assert np.array_equiv(v0_nan_mask, np.isnan(v1)), (v0, v1)
-            v0[v0_nan_mask] = 0
-            v1[v0_nan_mask] = 0
-        # check inf and rm it
-        v0_inf_mask = v0 == float("inf")
-        if np.any(v0_inf_mask):
-            assert np.array_equiv(v0_inf_mask, v1 == float("inf")), (v0, v1)
-            v0[v0_inf_mask] = 0
-            v1[v0_inf_mask] = 0
-        # check -inf and rm it
-        v0_inf_mask = v0 == float("-inf")
-        if np.any(v0_inf_mask):
-            assert np.array_equiv(v0_inf_mask, v1 == float("-inf")), (v0, v1)
-            v0[v0_inf_mask] = 0
-            v1[v0_inf_mask] = 0
-    else:
-        assert np.isfinite(v0.sum()) and np.isfinite(v1.sum()), (v0, v1)
-
-    assert v0.shape == v1.shape, "Two tensor must have same shape({} v.s. {})".format(
-        v0.shape, v1.shape
-    )
-    vdiv = np.max([np.abs(v0), np.abs(v1), np.ones_like(v0)], axis=0)
-    err = np.abs(v0 - v1) / vdiv
-    check = err > max_err
-    if check.sum():
-        idx = tuple(i[0] for i in np.nonzero(check))
-        if name is None:
-            name = "tensor"
-        else:
-            name = "tensor {}".format(name)
-        raise AssertionError(
-            "{} not equal: "
-            "shape={} nonequal_idx={} v0={} v1={} err={}".format(
-                name, v0.shape, idx, v0[idx], v1[idx], err[idx]
-            )
-        )
diff --git a/imperative/python/test/helpers/utils.py b/imperative/python/test/helpers/utils.py
index e62965826b5316ad5608a2d904e7fe338411fe6d..4724fd26a296977f0595df97839204bbb7b3596d 100644
--- a/imperative/python/test/helpers/utils.py
+++ b/imperative/python/test/helpers/utils.py
@@ -14,7 +14,8 @@ def opr_test(cases, func, compare_fn=_default_compare_fn, ref_fn=None, **kwargs)
            and should have output if ref_fn is None.
            should use list for multiple inputs and outputs for each case.
     :param func: the function to run opr.
-    :param compare_fn: the function to compare the result and expected, use assertTensorClose if None.
+    :param compare_fn: the function to compare the result and expected, use
+        ``np.testing.assert_allclose`` if None.
     :param ref_fn: the function to generate expected data, should assign output if None.
 
     Examples:
diff --git a/imperative/python/test/integration/test_correctness.py b/imperative/python/test/integration/test_correctness.py
index db77e374d28fc4d917e75b572fcb04223d5c9432..67e69dcfe6a1baf7cb9195615e3faa32196e52bc 100644
--- a/imperative/python/test/integration/test_correctness.py
+++ b/imperative/python/test/integration/test_correctness.py
@@ -24,7 +24,6 @@ from megengine.jit import SublinearMemoryConfig
 from megengine.module import AvgPool2d, BatchNorm2d, Conv2d, Linear, Module
 from megengine.optimizer import SGD
 from megengine.tensor import Tensor
-from megengine.test import assertTensorClose
 
 
 def get_gpu_name():
@@ -172,13 +171,13 @@ def run_train(
     loss = train_func(data, label, net, opt, gm)
     opt.step()
 
-    assertTensorClose(loss.numpy(), checkpoint["loss"], max_err=max_err)
+    np.testing.assert_allclose(loss.numpy(), checkpoint["loss"], atol=max_err)
 
     for param, param_ref in zip(
         net.state_dict().items(), checkpoint["net_updated"].items()
     ):
         assert param[0] == param_ref[0]
-        assertTensorClose(param[1], param_ref[1], max_err=max_err)
+        np.testing.assert_allclose(param[1], param_ref[1], atol=max_err)
 
 
 def run_eval(
@@ -209,7 +208,7 @@ def run_eval(
 
     for _ in range(3):
         new_value = eval_fun(data, net=net)
-        assertTensorClose(new_value.numpy(), refer_value.numpy(), max_err=max_err)
+        np.testing.assert_allclose(new_value.numpy(), refer_value.numpy(), atol=max_err)
 
 
 def test_correctness():
diff --git a/imperative/python/test/integration/test_dp_correctness.py b/imperative/python/test/integration/test_dp_correctness.py
index e35a4e8ffae6271e34876ccb6fd6d7f5fa5f646c..fcd8eaf2d41aa9419f985e694fa846e6e785b3cb 100644
--- a/imperative/python/test/integration/test_dp_correctness.py
+++ b/imperative/python/test/integration/test_dp_correctness.py
@@ -27,7 +27,6 @@ from megengine.functional.debug_param import set_conv_execution_strategy
 from megengine.module import AvgPool2d, BatchNorm2d, Conv2d, Linear, Module
 from megengine.optimizer import SGD
 from megengine.tensor import Tensor
-from megengine.test import assertTensorClose
 
 p_num = 4
 
@@ -181,7 +180,7 @@ def run_test(
 
         loss = train(data_train, label_train, net, opt, gm)
 
-        assertTensorClose(loss.numpy(), checkpoint["loss"], max_err=max_err)
+        np.testing.assert_allclose(loss.numpy(), checkpoint["loss"], atol=max_err)
 
         if dist.get_rank():
             return
@@ -189,7 +188,7 @@ def run_test(
             net.state_dict().items(), checkpoint["net_updated"].items()
         ):
             assert param[0] == param_ref[0]
-            assertTensorClose(param[1], param_ref[1], max_err=max_err)
+            np.testing.assert_allclose(param[1], param_ref[1], atol=max_err)
 
     procs = []
     for rank in range(p_num):
diff --git a/imperative/python/test/unit/core/test_dtype_quant.py b/imperative/python/test/unit/core/test_dtype_quant.py
index 78a3f24b9237c38486daece690f22f5674093d5b..e06cb297e016dce253ae910f4c1de39960234f60 100644
--- a/imperative/python/test/unit/core/test_dtype_quant.py
+++ b/imperative/python/test/unit/core/test_dtype_quant.py
@@ -1,4 +1,3 @@
-# -*- coding: utf-8 -*-
 # MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 #
 # Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
diff --git a/imperative/python/test/unit/core/test_megbrain_graph.py b/imperative/python/test/unit/core/test_megbrain_graph.py
index 3fb6a9de921947bfeceaec9dbc7c8ff993c04d4d..0abd92b331dcda747d51186f9fdf5c1c6c6dd888 100644
--- a/imperative/python/test/unit/core/test_megbrain_graph.py
+++ b/imperative/python/test/unit/core/test_megbrain_graph.py
@@ -11,7 +11,6 @@ from concurrent.futures import Future
 import numpy as np
 
 import megengine.functional as F
-from megengine.core._imperative_rt import DeviceTensorND
 from megengine.core.tensor import megbrain_graph as mgb_graph
 from megengine.core.tensor.raw_tensor import as_raw_tensor
 
diff --git a/imperative/python/test/unit/functional/test_elemwise.py b/imperative/python/test/unit/functional/test_elemwise.py
index db7db5cc09d355959f78329713f8c15d874cd850..0b167942929c89f3b6e5080b16cb46e0fb7fd6f2 100644
--- a/imperative/python/test/unit/functional/test_elemwise.py
+++ b/imperative/python/test/unit/functional/test_elemwise.py
@@ -10,34 +10,33 @@ import numpy as np
 
 import megengine.functional as F
 from megengine import tensor
-from megengine.test import assertTensorClose
 
 
 def test_abs():
-    assertTensorClose(
+    np.testing.assert_allclose(
         F.abs(tensor([-3.0, -4.0, -5.0])).numpy(),
         np.abs(np.array([-3.0, -4.0, -5.0], dtype=np.float32)),
     )
 
-    assertTensorClose(F.abs(-3.0).numpy(), np.abs(np.float32(-3.0)))
+    np.testing.assert_allclose(F.abs(-3.0).numpy(), np.abs(np.float32(-3.0)))
 
 
 def test_multiply():
-    assertTensorClose(
+    np.testing.assert_allclose(
         F.mul(-3.0, -4.0).numpy(), np.multiply(np.float32(-3.0), np.float32(-4.0))
     )
 
-    assertTensorClose(
+    np.testing.assert_allclose(
         F.mul(tensor([3.0, 4.0]), 4.0).numpy(),
         np.multiply(np.array([3.0, 4.0], dtype=np.float32), 4.0),
     )
 
-    assertTensorClose(
+    np.testing.assert_allclose(
         F.mul(4.0, tensor([3.0, 4.0])).numpy(),
         np.multiply(4.0, np.array([3.0, 4.0], dtype=np.float32)),
     )
 
-    assertTensorClose(
+    np.testing.assert_allclose(
         F.mul(tensor([3.0, 4.0]), tensor([3.0, 4.0])).numpy(),
         np.multiply(
             np.array([3.0, 4.0], dtype=np.float32),
@@ -51,24 +50,28 @@ def test_clamp():
     `F.clamp` will fall into wrong conditions unexpectedly.
     """
     x = np.linspace(-6, 6, dtype="float32")
-    assertTensorClose(F.clamp(tensor(x) + 3, 0, 6).numpy(), np.clip(x + 3, 0, 6))
-    assertTensorClose(F.clamp(tensor(x) - 3, -6, 0).numpy(), np.clip(x - 3, -6, 0))
+    np.testing.assert_allclose(
+        F.clamp(tensor(x) + 3, 0, 6).numpy(), np.clip(x + 3, 0, 6)
+    )
+    np.testing.assert_allclose(
+        F.clamp(tensor(x) - 3, -6, 0).numpy(), np.clip(x - 3, -6, 0)
+    )
 
 
 def test_isnan():
     for case in [[1, float("nan"), 0]]:
-        assertTensorClose(F.isnan(tensor(case)).numpy(), np.isnan(case))
+        np.testing.assert_allclose(F.isnan(tensor(case)).numpy(), np.isnan(case))
 
 
 def test_isinf():
     for case in [[1, float("inf"), 0]]:
-        assertTensorClose(F.isinf(tensor(case)).numpy(), np.isinf(case))
+        np.testing.assert_allclose(F.isinf(tensor(case)).numpy(), np.isinf(case))
 
 
 def test_sign():
     for case in [[1, -1, 0]]:
         x = tensor(case)
-        assertTensorClose(F.sign(x).numpy(), np.sign(case).astype(x.dtype))
+        np.testing.assert_allclose(F.sign(x).numpy(), np.sign(case).astype(x.dtype))
 
 
 def test_cosh():
diff --git a/imperative/python/test/unit/functional/test_functional.py b/imperative/python/test/unit/functional/test_functional.py
index 509b752a7cb0ad5852d0bd50e08c2f4dc347271b..22a5d7142963e4455f18082b2ae9b9a1aae7cbd2 100644
--- a/imperative/python/test/unit/functional/test_functional.py
+++ b/imperative/python/test/unit/functional/test_functional.py
@@ -19,7 +19,6 @@ from megengine import Parameter, Tensor, is_cuda_available, tensor
 from megengine.core._trace_option import use_tensor_shape
 from megengine.core.autodiff.grad import Grad
 from megengine.core.tensor.utils import make_shape_tuple
-from megengine.test import assertTensorClose
 
 
 def test_where():
@@ -105,10 +104,10 @@ def test_interpolate():
         out = F.interpolate(inp, scale_factor=2.0, mode="LINEAR")
         out2 = F.interpolate(inp, 4, mode="LINEAR")
 
-        assertTensorClose(
+        np.testing.assert_allclose(
             out.numpy(), np.array([[[1.0, 1.25, 1.75, 2.0]]], dtype=np.float32)
         )
-        assertTensorClose(
+        np.testing.assert_allclose(
             out2.numpy(), np.array([[[1.0, 1.25, 1.75, 2.0]]], dtype=np.float32)
         )
 
@@ -118,7 +117,7 @@ def test_interpolate():
         out = F.interpolate(inp, [4, 4])
         out2 = F.interpolate(inp, scale_factor=2.0)
 
-        assertTensorClose(out.numpy(), out2.numpy())
+        np.testing.assert_allclose(out.numpy(), out2.numpy())
 
     def assign_corner_interpolate():
         inp = tensor(np.arange(1, 5, dtype=np.float32).reshape(1, 1, 2, 2))
@@ -126,7 +125,7 @@ def test_interpolate():
         out = F.interpolate(inp, [4, 4], align_corners=True)
         out2 = F.interpolate(inp, scale_factor=2.0, align_corners=True)
 
-        assertTensorClose(out.numpy(), out2.numpy())
+        np.testing.assert_allclose(out.numpy(), out2.numpy())
 
     def error_shape_linear_interpolate():
         inp = tensor(np.arange(1, 5, dtype=np.float32).reshape(1, 1, 2, 2))
@@ -212,7 +211,7 @@ def test_one_hot():
         inp = tensor(np.arange(1, 4, dtype=np.int32))
         out = F.one_hot(inp, num_classes=4)
 
-        assertTensorClose(
+        np.testing.assert_allclose(
             out.numpy(), np.eye(4, dtype=np.int32)[np.arange(1, 4, dtype=np.int32)]
         )
 
@@ -225,7 +224,7 @@ def test_one_hot():
         inp = tensor(arr)
         out = F.one_hot(inp, 10)
 
-        assertTensorClose(out.numpy(), np.eye(10, dtype=np.int32)[arr])
+        np.testing.assert_allclose(out.numpy(), np.eye(10, dtype=np.int32)[arr])
 
     onehot_low_dimension()
     onehot_high_dimension()
@@ -237,16 +236,16 @@ def test_add_update():
     b = Tensor(v)
 
     u = F.add_update(b, 1)
-    assertTensorClose(u.numpy(), v + 1)
+    np.testing.assert_allclose(u.numpy(), v + 1, atol=1e-6)
     u = F.add_update(b, 1)
-    assertTensorClose(u.numpy(), v + 2)
+    np.testing.assert_allclose(u.numpy(), v + 2, atol=1e-6)
 
     x = np.ones((2, 2), dtype=np.float32)
     y = x * 0.5
     dest = tensor(x)
     delta = tensor(y)
     r = F.add_update(dest, delta, alpha=0.9, beta=0.1, bias=0.1)
-    assertTensorClose(r.numpy(), x * 0.9 + y * 0.1 + 0.1)
+    np.testing.assert_allclose(r.numpy(), x * 0.9 + y * 0.1 + 0.1, atol=1e-6)
 
 
 def test_add_update_params():
@@ -263,7 +262,7 @@ def test_add_update_params():
     F.add_update(y, z, beta=0.1)
 
     res = f(np.ones((2, 3)).astype(np.float32))
-    assertTensorClose(res.numpy(), b + 1)
+    np.testing.assert_allclose(res.numpy(), b + 1)
 
 
 def test_binary_cross_entropy():
@@ -276,7 +275,7 @@ def test_binary_cross_entropy():
         return 1 / (1 + np.exp(-x))
 
     def compare_fn(x, y):
-        assertTensorClose(x.numpy(), y, max_err=5e-4)
+        np.testing.assert_allclose(x.numpy(), y, atol=5e-4)
 
     np.random.seed(123)
     data1 = sigmoid(np.random.uniform(size=data1_shape).astype(np.float32))
@@ -444,7 +443,7 @@ def test_conv_bias():
             result = F.transpose(result, (0, 1, 4, 2, 3))
         expected = F.flatten(expected)
         result = F.flatten(result)
-        assertTensorClose(result.numpy(), expected.numpy(), max_err=outp_scale)
+        np.testing.assert_allclose(result.numpy(), expected.numpy(), atol=outp_scale)
 
     run(1, 4, 4, 24, 33, 1, 1, 2, 3, 1, 1, False)
     run(10, 12, 24, 46, 46, 1, 1, 2, 1, 3, 1, False)
diff --git a/imperative/python/test/unit/functional/test_math.py b/imperative/python/test/unit/functional/test_math.py
index e0c28a3b5111844ea5da129df6f91a7beddf62b2..aec0366a169a7c112a9a215174ce0e2ccc6ca270 100644
--- a/imperative/python/test/unit/functional/test_math.py
+++ b/imperative/python/test/unit/functional/test_math.py
@@ -13,7 +13,6 @@ from utils import opr_test
 
 import megengine.functional as F
 from megengine import tensor
-from megengine.test import assertTensorClose
 
 
 def common_test_reduce(opr, ref_opr):
diff --git a/imperative/python/test/unit/functional/test_tensor.py b/imperative/python/test/unit/functional/test_tensor.py
index 38193ff31f6f241c989eee8f787c47aec4a4e6f7..72da5d861a14b902ea8eead601ebc6a496ee5c64 100644
--- a/imperative/python/test/unit/functional/test_tensor.py
+++ b/imperative/python/test/unit/functional/test_tensor.py
@@ -18,14 +18,13 @@ from megengine import tensor
 from megengine.core._trace_option import use_tensor_shape
 from megengine.core.tensor.utils import astensor1d
 from megengine.distributed.helper import get_device_count_by_fork
-from megengine.test import assertTensorClose
 
 
 def test_eye():
     dtype = np.float32
     cases = [{"input": [10, 20]}, {"input": [20, 30]}]
     for case in cases:
-        assertTensorClose(
+        np.testing.assert_allclose(
             F.eye(case["input"], dtype=dtype).numpy(),
             np.eye(*case["input"]).astype(dtype),
         )
diff --git a/imperative/python/test/unit/module/test_activation.py b/imperative/python/test/unit/module/test_activation.py
index afca1fde0cf5781d33b580d808d3b86f0fa3bcf7..8212b47e81a98b264d7968668bfc1710778a87ab 100644
--- a/imperative/python/test/unit/module/test_activation.py
+++ b/imperative/python/test/unit/module/test_activation.py
@@ -10,7 +10,6 @@ import numpy as np
 
 import megengine as mge
 from megengine.module import LeakyReLU
-from megengine.test import assertTensorClose
 
 
 def test_leaky_relu():
@@ -21,4 +20,4 @@ def test_leaky_relu():
     output = leaky_relu(mge.tensor(data))
 
     np_output = np.maximum(0, data) + negative_slope * np.minimum(0, data)
-    assertTensorClose(output.numpy(), np_output, max_err=0)
+    np.testing.assert_equal(output.numpy(), np_output)
diff --git a/imperative/python/test/unit/module/test_batchnorm.py b/imperative/python/test/unit/module/test_batchnorm.py
index c99dd6b34a4e1b3ea67a32e85f4bb9fc05360f7f..d3debb12dd9d219d3fcea6d01169ed4c01771680 100644
--- a/imperative/python/test/unit/module/test_batchnorm.py
+++ b/imperative/python/test/unit/module/test_batchnorm.py
@@ -17,7 +17,6 @@ import megengine.distributed as dist
 from megengine import Tensor
 from megengine.core._trace_option import use_tensor_shape
 from megengine.module import BatchNorm1d, BatchNorm2d, SyncBatchNorm
-from megengine.test import assertTensorClose
 
 
 @pytest.mark.skipif(
@@ -47,9 +46,9 @@ def test_syncbn():
         for i in range(steps):
             yv = bn(Tensor(data[i]))
 
-        assertTensorClose(yv_expect, yv.numpy(), max_err=5e-6)
-        assertTensorClose(running_mean, bn.running_mean.numpy(), max_err=5e-6)
-        assertTensorClose(running_var, bn.running_var.numpy(), max_err=5e-6)
+        np.testing.assert_allclose(yv.numpy(), yv_expect, atol=5e-6)
+        np.testing.assert_allclose(bn.running_mean.numpy(), running_mean, atol=5e-6)
+        np.testing.assert_allclose(bn.running_var.numpy(), running_var, atol=5e-6)
 
     xv = []
     for i in range(steps):
@@ -119,12 +118,12 @@ def test_batchnorm():
         yv = bn(Tensor(xv))
         yv_expect = (xv - mean) / sd
 
-        assertTensorClose(yv_expect, yv.numpy(), max_err=5e-6)
-        assertTensorClose(
-            running_mean.reshape(-1), bn.running_mean.numpy().reshape(-1), max_err=5e-6
+        np.testing.assert_allclose(yv.numpy(), yv_expect, atol=5e-6)
+        np.testing.assert_allclose(
+            bn.running_mean.numpy().reshape(-1), running_mean.reshape(-1), atol=5e-6
         )
-        assertTensorClose(
-            running_var.reshape(-1), bn.running_var.numpy().reshape(-1), max_err=5e-6
+        np.testing.assert_allclose(
+            bn.running_var.numpy().reshape(-1), running_var.reshape(-1), atol=5e-6
         )
 
     # test set 'training' flag to False
@@ -135,11 +134,11 @@ def test_batchnorm():
     data = Tensor(xv)
     yv1 = bn(data)
     yv2 = bn(data)
-    assertTensorClose(yv1.numpy(), yv2.numpy(), max_err=0)
-    assertTensorClose(mean_backup, bn.running_mean.numpy(), max_err=0)
-    assertTensorClose(var_backup, bn.running_var.numpy(), max_err=0)
+    np.testing.assert_equal(yv1.numpy(), yv2.numpy())
+    np.testing.assert_equal(mean_backup, bn.running_mean.numpy())
+    np.testing.assert_equal(var_backup, bn.running_var.numpy())
     yv_expect = (xv - running_mean) / np.sqrt(running_var + bn.eps)
-    assertTensorClose(yv_expect, yv1.numpy(), max_err=5e-6)
+    np.testing.assert_allclose(yv1.numpy(), yv_expect, atol=5e-6)
 
 
 @pytest.mark.skipif(
@@ -173,12 +172,12 @@ def test_syncbn1d():
         yv = bn(Tensor(xv))
         yv_expect = (xv - mean) / sd
 
-        assertTensorClose(yv_expect, yv.numpy(), max_err=5e-6)
-        assertTensorClose(
-            running_mean.reshape(-1), bn.running_mean.numpy().reshape(-1), max_err=5e-6
+        np.testing.assert_allclose(yv.numpy(), yv_expect, atol=5e-6)
+        np.testing.assert_allclose(
+            bn.running_mean.numpy().reshape(-1), running_mean.reshape(-1), atol=5e-6
         )
-        assertTensorClose(
-            running_var.reshape(-1), bn.running_var.numpy().reshape(-1), max_err=5e-6
+        np.testing.assert_allclose(
+            bn.running_var.numpy().reshape(-1), running_var.reshape(-1), atol=5e-6
         )
 
     # test set 'training' flag to False
@@ -189,11 +188,11 @@ def test_syncbn1d():
     data = Tensor(xv)
     yv1 = bn(data)
     yv2 = bn(data)
-    assertTensorClose(yv1.numpy(), yv2.numpy(), max_err=0)
-    assertTensorClose(mean_backup, bn.running_mean.numpy(), max_err=0)
-    assertTensorClose(var_backup, bn.running_var.numpy(), max_err=0)
+    np.testing.assert_equal(yv1.numpy(), yv2.numpy())
+    np.testing.assert_equal(mean_backup, bn.running_mean.numpy())
+    np.testing.assert_equal(var_backup, bn.running_var.numpy())
     yv_expect = (xv - running_mean) / np.sqrt(running_var + bn.eps)
-    assertTensorClose(yv_expect, yv1.numpy(), max_err=5e-6)
+    np.testing.assert_allclose(yv1.numpy(), yv_expect, atol=5e-6)
 
 
 def test_batchnorm2d():
@@ -221,9 +220,9 @@ def test_batchnorm2d():
         yv = bn(Tensor(xv))
         yv_expect = (xv - mean) / sd
 
-        assertTensorClose(yv_expect, yv.numpy(), max_err=5e-6)
-        assertTensorClose(running_mean, bn.running_mean.numpy(), max_err=5e-6)
-        assertTensorClose(running_var, bn.running_var.numpy(), max_err=5e-6)
+        np.testing.assert_allclose(yv.numpy(), yv_expect, atol=5e-6)
+        np.testing.assert_allclose(bn.running_mean.numpy(), running_mean, atol=5e-6)
+        np.testing.assert_allclose(bn.running_var.numpy(), running_var, atol=5e-6)
 
     # test set 'training' flag to False
     mean_backup = bn.running_mean.numpy()
@@ -233,11 +232,11 @@ def test_batchnorm2d():
     data = Tensor(xv)
     yv1 = bn(data)
     yv2 = bn(data)
-    assertTensorClose(yv1.numpy(), yv2.numpy(), max_err=0)
-    assertTensorClose(mean_backup, bn.running_mean.numpy(), max_err=0)
-    assertTensorClose(var_backup, bn.running_var.numpy(), max_err=0)
+    np.testing.assert_equal(yv1.numpy(), yv2.numpy())
+    np.testing.assert_equal(mean_backup, bn.running_mean.numpy())
+    np.testing.assert_equal(var_backup, bn.running_var.numpy())
     yv_expect = (xv - running_mean) / np.sqrt(running_var + bn.eps)
-    assertTensorClose(yv_expect, yv1.numpy(), max_err=5e-6)
+    np.testing.assert_allclose(yv1.numpy(), yv_expect, atol=5e-6)
 
 
 @pytest.mark.skipif(
@@ -272,9 +271,9 @@ def test_syncbn2d():
         yv = bn(Tensor(xv))
         yv_expect = (xv - mean) / sd
 
-        assertTensorClose(yv_expect, yv.numpy(), max_err=5e-6)
-        assertTensorClose(running_mean, bn.running_mean.numpy(), max_err=5e-6)
-        assertTensorClose(running_var, bn.running_var.numpy(), max_err=5e-6)
+        np.testing.assert_allclose(yv.numpy(), yv_expect, atol=5e-6)
+        np.testing.assert_allclose(bn.running_mean.numpy(), running_mean, atol=5e-6)
+        np.testing.assert_allclose(bn.running_var.numpy(), running_var, atol=5e-6)
 
     # test set 'training' flag to False
     mean_backup = bn.running_mean.numpy()
@@ -284,11 +283,11 @@ def test_syncbn2d():
     data = Tensor(xv)
     yv1 = bn(data)
     yv2 = bn(data)
-    assertTensorClose(yv1.numpy(), yv2.numpy(), max_err=0)
-    assertTensorClose(mean_backup, bn.running_mean.numpy(), max_err=0)
-    assertTensorClose(var_backup, bn.running_var.numpy(), max_err=0)
+    np.testing.assert_equal(yv1.numpy(), yv2.numpy())
+    np.testing.assert_equal(mean_backup, bn.running_mean.numpy())
+    np.testing.assert_equal(var_backup, bn.running_var.numpy())
     yv_expect = (xv - running_mean) / np.sqrt(running_var + bn.eps)
-    assertTensorClose(yv_expect, yv1.numpy(), max_err=5e-6)
+    np.testing.assert_allclose(yv1.numpy(), yv_expect, atol=5e-6)
 
 
 def test_batchnorm_no_stats():
@@ -311,7 +310,7 @@ def test_batchnorm_no_stats():
         yv = bn(Tensor(xv))
         yv_expect = (xv - mean) / sd
 
-        assertTensorClose(yv_expect, yv.numpy(), max_err=5e-6)
+        np.testing.assert_allclose(yv.numpy(), yv_expect, atol=5e-6)
 
 
 @pytest.mark.skipif(
@@ -341,7 +340,7 @@ def test_syncbn_no_stats():
         yv = bn(Tensor(xv))
         yv_expect = (xv - mean) / sd
 
-        assertTensorClose(yv_expect, yv.numpy(), max_err=5e-6)
+        np.testing.assert_allclose(yv.numpy(), yv_expect, atol=5e-6)
 
 
 def test_batchnorm2d_no_stats():
@@ -363,7 +362,7 @@ def test_batchnorm2d_no_stats():
         yv = bn(Tensor(xv))
         yv_expect = (xv - mean) / sd
 
-        assertTensorClose(yv_expect, yv.numpy(), max_err=5e-6)
+        np.testing.assert_allclose(yv.numpy(), yv_expect, atol=5e-6)
 
 
 @pytest.mark.skipif(
@@ -392,4 +391,4 @@ def test_syncbn2d_no_stats():
         yv = bn(Tensor(xv))
         yv_expect = (xv - mean) / sd
 
-        assertTensorClose(yv_expect, yv.numpy(), max_err=5e-6)
+        np.testing.assert_allclose(yv.numpy(), yv_expect, atol=5e-6)
diff --git a/imperative/python/test/unit/module/test_conv.py b/imperative/python/test/unit/module/test_conv.py
index 262498f9475330d3ddf4fa4f659da247b5c17def..7fdbdbe8156c088fa085217a361d7e19dae97e84 100644
--- a/imperative/python/test/unit/module/test_conv.py
+++ b/imperative/python/test/unit/module/test_conv.py
@@ -12,7 +12,6 @@ import numpy as np
 
 from megengine import Parameter, tensor
 from megengine.module import ConvTranspose2d, LocalConv2d
-from megengine.test import assertTensorClose
 
 
 def test_conv_transpose2d():
@@ -49,7 +48,7 @@ def test_conv_transpose2d():
         conv_transpose2d.bias = Parameter(bias, dtype=np.float32)
     y = conv_transpose2d(tensor(inp))
 
-    assertTensorClose(out, y.numpy(), max_err=2e-6)
+    np.testing.assert_allclose(out, y.numpy(), atol=2e-6)
 
 
 def test_local_conv2d():
@@ -107,4 +106,4 @@ def test_local_conv2d():
             * weights[0, oh, ow, :, :, :, oc]
         )
 
-    assertTensorClose(outputs.numpy(), expected, max_err=1e-5)
+    np.testing.assert_allclose(outputs.numpy(), expected, atol=1e-5)
diff --git a/imperative/python/test/unit/module/test_module.py b/imperative/python/test/unit/module/test_module.py
index 48c815b135facebd226fcaf097ea863808bcbde7..333498d0d54f58493ce00490851d29106d54e272 100644
--- a/imperative/python/test/unit/module/test_module.py
+++ b/imperative/python/test/unit/module/test_module.py
@@ -29,7 +29,6 @@ from megengine.module import (
     Softmax,
 )
 from megengine.quantization.quantize import quantize, quantize_qat
-from megengine.test import assertTensorClose
 
 
 class MLP(Module):
@@ -87,7 +86,7 @@ def graph_mode(*modes):
 
 
 def _default_compare_fn(x, y):
-    assertTensorClose(x.numpy(), y)
+    np.testing.assert_allclose(x.numpy(), y, rtol=1e-6)
 
 
 def opr_test(
@@ -102,7 +101,7 @@ def opr_test(
     mode: the list of test mode which are eager, static and dynamic_shape
           will test all the cases if None.
     func: the function to run opr.
-    compare_fn: the function to compare the result and expected, use assertTensorClose if None.
+    compare_fn: the function to compare the result and expected, use np.testing.assert_allclose if None.
     ref_fn: the function to generate expected data, should assign output if None.
     cases: the list which have dict element, the list length should be 2 for dynamic shape test.
            and the dict should have input,
@@ -331,17 +330,17 @@ def test_module_api_hooks():
     bn1 = F.batch_norm2d(
         x + 3, mean1, Parameter(np.ones(shape), dtype=np.float32), training=True
     )
-    assertTensorClose(
+    np.testing.assert_allclose(
         net.i.bn.running_mean.numpy(), mean1.numpy(),
     )
     mean2 = Parameter(np.zeros(shape), dtype=np.float32)
     bn2 = F.batch_norm2d(
         bn1 + 3, mean2, Parameter(np.ones(shape), dtype=np.float32), training=True
     )
-    assertTensorClose(
+    np.testing.assert_allclose(
         net.bn.running_mean.numpy(), mean2.numpy(),
     )
-    assertTensorClose((bn2 + 2).numpy(), y.numpy())
+    np.testing.assert_allclose((bn2 + 2).numpy(), y.numpy())
 
     assert len(hooks) == 8
     for handler in hooks:
@@ -479,7 +478,7 @@ def test_state_dict():
         mlp1 = MLP()
         mlp1.load_state_dict(state_dict, strict=False)
         pred1 = mlp1(data)
-        assertTensorClose(pred0.numpy(), pred1.numpy(), max_err=5e-6)
+        np.testing.assert_allclose(pred0.numpy(), pred1.numpy(), atol=5e-6)
         with pytest.raises(KeyError):
             mlp1.load_state_dict(state_dict)
         del state_dict["extra"]
@@ -520,13 +519,13 @@ def test_shared_param():
     net = Simple()
     assert net.conv0.weight is net.conv1.weight
     data = tensor(np.random.random((1, 1, 8, 8)).astype(np.float32))
-    assertTensorClose(net.conv0(data).numpy(), net.conv1(data).numpy())
+    np.testing.assert_allclose(net.conv0(data).numpy(), net.conv1(data).numpy())
     with BytesIO() as f:
         mge.save(net, f)
         f.seek(0)
         net1 = mge.load(f)
     assert net1.conv0.weight is net1.conv1.weight
-    assertTensorClose(net1.conv0(data).numpy(), net1.conv1(data).numpy())
+    np.testing.assert_allclose(net1.conv0(data).numpy(), net1.conv1(data).numpy())
 
     with BytesIO() as f:
         mge.save(net.conv0, f)
@@ -539,7 +538,7 @@ def test_shared_param():
         conv1 = mge.load(f)
 
     assert conv0.weight is not conv1.weight
-    assertTensorClose(conv0(data).numpy(), conv1(data).numpy())
+    np.testing.assert_allclose(conv0(data).numpy(), conv1(data).numpy())
 
 
 def test_pickle_module():
@@ -562,8 +561,8 @@ def test_pickle_module():
         mlp1 = mge.load(fout)
         pred2 = mlp1(data)
 
-    assertTensorClose(pred0.numpy(), pred1.numpy(), max_err=5e-6)
-    assertTensorClose(pred0.numpy(), pred2.numpy(), max_err=5e-6)
+    np.testing.assert_allclose(pred0.numpy(), pred1.numpy(), atol=5e-6)
+    np.testing.assert_allclose(pred0.numpy(), pred2.numpy(), atol=5e-6)
 
 
 @pytest.mark.skip(reason="under development")
@@ -609,8 +608,8 @@ def test_load_quantized():
         mlp.load_state_dict(checkpoint)
         pred1 = mlp(data)
 
-    assertTensorClose(
-        pred0.astype("float32").numpy(), pred1.astype("float32").numpy(), max_err=5e-6
+    np.testing.assert_allclose(
+        pred0.astype("float32").numpy(), pred1.astype("float32").numpy(), atol=5e-6
     )
 
 
diff --git a/imperative/python/test/unit/module/test_module_tensor.py b/imperative/python/test/unit/module/test_module_tensor.py
index 38686b08c7b3ab18610e55268160d4b009538a29..773bf373a05ee1f76586f029783c3495fd0c337f 100644
--- a/imperative/python/test/unit/module/test_module_tensor.py
+++ b/imperative/python/test/unit/module/test_module_tensor.py
@@ -15,7 +15,6 @@ import megengine as mge
 import megengine.functional as F
 from megengine import Parameter, Tensor
 from megengine.module import Conv2d
-from megengine.test import assertTensorClose
 
 
 def test_set_value():
@@ -23,21 +22,21 @@ def test_set_value():
     param = Parameter(v0)
     v1 = np.random.random((2, 3)).astype(np.float32)
     param.set_value(v1)
-    assertTensorClose(param.numpy(), v1, max_err=5e-6)
+    np.testing.assert_allclose(param.numpy(), v1, atol=5e-6)
     v2 = np.random.random((3, 3)).astype(np.float32)
     # TODO: add this
     # with pytest.raises(ValueError):
     #     param.set_value(v2)
-    assertTensorClose(param.numpy(), v1, max_err=5e-6)
+    np.testing.assert_allclose(param.numpy(), v1, atol=5e-6)
 
 
 @pytest.mark.skip(reason="fill unsupported")
 def test_fill():
     a = Tensor(np.zeros((2, 3), dtype=np.float32))
     a.fill(3)
-    assertTensorClose(a.numpy(), np.full((2, 3), 3, dtype=np.float32))
+    np.testing.assert_allclose(a.numpy(), np.full((2, 3), 3, dtype=np.float32))
     a.fill(124.568)
-    assertTensorClose(a.numpy(), np.full((2, 3), 124.568, dtype=np.float32))
+    np.testing.assert_allclose(a.numpy(), np.full((2, 3), 124.568, dtype=np.float32))
 
 
 # TODO: remove or rewrite following test
@@ -51,11 +50,11 @@ def test_fill():
 #         f = compile(v, None)
 
 #     out, = f()
-#     assertTensorClose(out, p_ * 2)
+#     np.testing.assert_allclose(out, p_ * 2)
 
 #     F.add_update(p, p)
 #     out, = f()
-#     assertTensorClose(out, p_ * 4)
+#     np.testing.assert_allclose(out, p_ * 4)
 
 
 # TODO: remove or rewrite following test
@@ -74,7 +73,7 @@ def test_fill():
 #     data1 = Input("data", value=v)
 #     out1 = net(data1)
 
-#     assertTensorClose(out0, out1.numpy())
+#     np.testing.assert_allclose(out0, out1.numpy())
 
 
 # def test_shape_warning():
diff --git a/imperative/python/test/unit/module/test_qat.py b/imperative/python/test/unit/module/test_qat.py
index d9312bb492ef701d45a63ba1b15b1929686e7f63..718d6acbb7491e09b94b8581fe14632a5c86c2ea 100644
--- a/imperative/python/test/unit/module/test_qat.py
+++ b/imperative/python/test/unit/module/test_qat.py
@@ -12,7 +12,6 @@ from megengine.module import (
     QuantStub,
 )
 from megengine.quantization.quantize import disable_fake_quant, quantize_qat
-from megengine.test import assertTensorClose
 
 
 def test_qat_convbn2d():
@@ -31,22 +30,24 @@ def test_qat_convbn2d():
         # import pdb
         # pdb.set_trace()
         qat_outputs = qat_module(inputs)
-        assertTensorClose(normal_outputs.numpy(), qat_outputs.numpy(), max_err=5e-6)
-        assertTensorClose(
+        np.testing.assert_allclose(
+            normal_outputs.numpy(), qat_outputs.numpy(), atol=5e-6
+        )
+        np.testing.assert_allclose(
             module.bn.running_mean.numpy(),
             qat_module.bn.running_mean.numpy(),
-            max_err=5e-8,
+            atol=5e-8,
         )
-        assertTensorClose(
-            module.bn.running_var.numpy(),
-            qat_module.bn.running_var.numpy(),
-            max_err=5e-7,
+        np.testing.assert_allclose(
+            module.bn.running_var.numpy(), qat_module.bn.running_var.numpy(), atol=5e-7,
         )
         module.eval()
         normal_outputs = module(inputs)
         qat_module.eval()
         qat_outputs = qat_module(inputs)
-        assertTensorClose(normal_outputs.numpy(), qat_outputs.numpy(), max_err=5e-6)
+        np.testing.assert_allclose(
+            normal_outputs.numpy(), qat_outputs.numpy(), atol=5e-6
+        )
 
 
 def test_qat_conv():
@@ -82,10 +83,10 @@ def test_qat_conv():
         disable_fake_quant(qat_net)
         normal_outputs = net(inputs)
         qat_outputs = qat_net(inputs)
-        assertTensorClose(normal_outputs.numpy(), qat_outputs.numpy())
+        np.testing.assert_allclose(normal_outputs.numpy(), qat_outputs.numpy())
 
         net.eval()
         normal_outputs = net(inputs)
         qat_net.eval()
         qat_outputs = qat_net(inputs)
-        assertTensorClose(normal_outputs.numpy(), qat_outputs.numpy())
+        np.testing.assert_allclose(normal_outputs.numpy(), qat_outputs.numpy())
diff --git a/imperative/python/test/unit/quantization/test_fake_quant.py b/imperative/python/test/unit/quantization/test_fake_quant.py
index ff999b75828b2eb8dbc3d4ae41779c8a198b457f..dc82182fa6a227088cb5acdf7a70eb83f0713ea3 100644
--- a/imperative/python/test/unit/quantization/test_fake_quant.py
+++ b/imperative/python/test/unit/quantization/test_fake_quant.py
@@ -13,7 +13,6 @@ import megengine as mge
 from megengine import tensor
 from megengine.quantization.fake_quant import TQT_Function
 from megengine.quantization.internal_fake_quant import *
-from megengine.test import assertTensorClose
 
 
 class numpy_TQT_Function:
@@ -60,13 +59,13 @@ def test_TQT():
     nf = numpy_TQT_Function(-127, 127)
 
     def check_inp(a, b, c, a_np, b_np, c_np):
-        assertTensorClose(
-            f.forward(a, b).numpy(), nf.forward(a_np, b_np).astype("float32")
+        np.testing.assert_allclose(
+            f.forward(a, b).numpy(), nf.forward(a_np, b_np).astype("float32"), rtol=1e-6
         )
         c1, c2 = f.backward(c)
         c1_np, c2_np = nf.backward(c_np)
-        assertTensorClose(c1.numpy(), c1_np.astype("float32"))
-        assertTensorClose(c2.numpy(), c2_np.astype("float32"))
+        np.testing.assert_allclose(c1.numpy(), c1_np.astype("float32"), rtol=1e-6)
+        np.testing.assert_allclose(c2.numpy(), c2_np.astype("float32"), rtol=1e-6)
 
     a_np = np.random.random((4, 3)).astype("float32")
     b_np = np.random.random((1)).astype("float32")