test(mge/parampack): add parampack related opr test

GitOrigin-RevId: 550c6b6443c36ff7afcbaeae1f4c2d5fdc29c7d3

imperative/python/megengine/functional/__init__.py

@@ -25,7 +25,7 @@ from .math import *
 from .nn import *
 from .quantized import conv_bias_activation
 from .tensor import *
-from .utils import accuracy, zero_grad
+from .utils import accuracy, copy, zero_grad
 
 # delete namespace
 # pylint: disable=undefined-variable

imperative/python/megengine/functional/tensor.py

@@ -980,9 +980,9 @@ def param_pack_concat(inps: List, offsets: Tensor, offsets_val: List) -> Tensor:
     Returns concat Tensor, only used for parampack.
 
     :param inps: Input tensors
-    :param offsets: offsets of inputs, length of 2 * n,
+    :param offsets: device value of offsets
+    :param offsets_val: offsets of inputs, length of 2 * n,
             format [begin0, end0, begin1, end1].
-    :param offsets_val: device value of offsets
     :return: split tensors
 
     Examples:
@@ -995,8 +995,8 @@ def param_pack_concat(inps: List, offsets: Tensor, offsets_val: List) -> Tensor:
 
         a = tensor(np.ones((1,), np.int32))
         b = tensor(np.ones((3, 3), np.int32))
-        offsets = [0, 1, 1, 10]
-        offsets_val = tensor(offsets, np.int32)
+        offsets_val = [0, 1, 1, 10]
+        offsets = tensor(offsets, np.int32)
         c = F.param_pack_concat([a, b], offsets, offsets_val)
         print(c.numpy())
 

imperative/python/test/unit/functional/test_tensor.py

@@ -314,3 +314,42 @@ def test_device():
     y5 = F.full((3, 2), 4, device=x.device)
     y6 = F.full((3, 2), 4, device="xpux")
     np.testing.assert_almost_equal(y5.numpy(), y6.numpy())
+
+
+def copy_test(dst, src):
+    data = np.random.random((2, 3)).astype(np.float32)
+    x = tensor(data, device=src)
+    y = F.copy(x, dst)
+    assert np.allclose(data, y.numpy())
+
+
+@pytest.mark.skipif(not is_cuda_available(), reason="CUDA is disabled")
+def test_copy_h2d():
+    copy_test("cpu0", "gpu0")
+
+
+@pytest.mark.skipif(not is_cuda_available(), reason="CUDA is disabled")
+def test_copy_d2h():
+    copy_test("gpu0", "cpu0")
+
+
+@pytest.mark.skipif(not is_cuda_available(), reason="CUDA is disabled")
+def test_copy_d2d():
+    copy_test("gpu0", "gpu1")
+    copy_test("gpu0:0", "gpu0:1")
+
+
+def test_param_pack_split():
+    a = tensor(np.ones((10,), np.int32))
+    b, c = F.param_pack_split(a, [0, 1, 1, 10], [(1,), (3, 3)])
+    assert np.allclose(b.numpy(), a.numpy()[1])
+    assert np.allclose(c.numpy(), a.numpy()[1:].reshape(3, 3))
+
+
+def test_param_pack_concat():
+    a = tensor(np.ones((1,), np.int32))
+    b = tensor(np.ones((3, 3), np.int32))
+    offsets_val = [0, 1, 1, 10]
+    offsets = tensor(offsets_val, np.int32)
+    c = F.param_pack_concat([a, b], offsets, offsets_val)
+    assert np.allclose(np.concatenate([a.numpy(), b.numpy().flatten()]), c.numpy())