diff --git a/mindnlp/core/_tensor.py b/mindnlp/core/_tensor.py
index 7ee9ea8f7..c122e0b2d 100644
--- a/mindnlp/core/_tensor.py
+++ b/mindnlp/core/_tensor.py
@@ -812,6 +812,9 @@ def record_stream(self, stream):
     Tensor.record_stream = record_stream
     StubTensor.record_stream = record_stream
 
+    Tensor.scatter = ops.scatter
+    StubTensor.scatter = ops.scatter
+
 def _rebuild_from_type_v2(func, new_type, args, state):
     ret = func(*args)
     return ret
\ No newline at end of file
diff --git a/mindnlp/core/configs.py b/mindnlp/core/configs.py
index 6543b55c1..4711dfcc8 100644
--- a/mindnlp/core/configs.py
+++ b/mindnlp/core/configs.py
@@ -4,7 +4,7 @@
 
 SOC = MSContext.get_instance().get_ascend_soc_version()
 DEVICE_TARGET = mindspore.get_context('device_target')
-SUPPORT_BF16 = DEVICE_TARGET == 'Ascend' and SOC not in ['ascend910', 'ascend310b1', 'ascend310b4']
+SUPPORT_BF16 = DEVICE_TARGET == 'Ascend' and SOC not in ['ascend910', 'ascend310b']
 ON_A1 = not SUPPORT_BF16
 ON_ORANGE_PI = '310b' in SOC
 USE_PYBOOST = DEVICE_TARGET == 'Ascend'
diff --git a/mindnlp/core/npu/__init__.py b/mindnlp/core/npu/__init__.py
index c5ded5b1f..febdda27c 100644
--- a/mindnlp/core/npu/__init__.py
+++ b/mindnlp/core/npu/__init__.py
@@ -2,9 +2,14 @@
 
 import mindspore
 from mindspore import get_rng_state, set_rng_state, manual_seed
-from mindspore.hal import *
+from mindspore.runtime import memory_reserved as ms_memory_reserved, \
+    memory_allocated as ms_memory_allocated, StreamCtx as StreamContext, Stream, empty_cache, \
+    reset_peak_memory_stats, reset_max_memory_allocated, max_memory_allocated, synchronize, \
+    current_stream
+from mindspore.device_context.ascend import device_count 
 
 from mindnlp import core
+from ..configs import SUPPORT_BF16
 
 FloatTensor = core.FloatTensor
 HalfTensor = core.FloatTensor
@@ -28,6 +33,15 @@ def set_device(device):
 def _lazy_call(callable, **kwargs):
     callable()
 
+def is_bf16_supported():
+    return SUPPORT_BF16
+
+def memory_allocated(device=None):
+    return ms_memory_allocated()
+
+def memory_reserved(device=None):
+    return ms_memory_reserved()
+
 class device:
     r"""Context-manager that changes the selected device.
 
diff --git a/mindnlp/core/ops/array.py b/mindnlp/core/ops/array.py
index d1ed73f28..f01a9ab90 100644
--- a/mindnlp/core/ops/array.py
+++ b/mindnlp/core/ops/array.py
@@ -71,11 +71,13 @@ def gather(input, dim, index):
         _complex = _get_cache_prim(ops.Complex)()
         return _complex(real_part, imag_part)
 
-    if use_pyboost() and has_gather:
+    if use_pyboost() and has_gather and not ON_ORANGE_PI:
         return mindspore.mint.gather(input, dim, index)
 
-    index = ops.where(index < input.shape[dim], index, index - input.shape[dim])
-    return ops.gather_elements(input, dim, index)
+    index = core.where(index < input.shape[dim], index, index - input.shape[dim])
+    if not ON_ORANGE_PI:
+        return ops.gather_elements(input, dim, index)
+    return tf_gather(input, index, dim, batch_dims=dim)
 
 def gather_nd(input, indices):
     return ops.gather_nd(input, indices)
@@ -195,10 +197,12 @@ def select(input, dim, index):
 # scatter
 has_scatter = hasattr(mindspore.mint, 'scatter')
 def scatter(input, dim, index, src):
-    if use_pyboost() and has_scatter:
+    if use_pyboost() and has_scatter and not ON_ORANGE_PI:
         return mindspore.mint.scatter(input, dim, index, src)
     if not isinstance(src, mindspore.Tensor):
         src = ops.full(index.shape, src, dtype=input.dtype)
+    if input.dtype == mindspore.bool_:
+        return ops.tensor_scatter_elements(input.int(), index, src.int(), dim).bool()
     return ops.tensor_scatter_elements(input, index, src, dim)
 
 def tf_scatter_nd_update(input, indices, updates):
@@ -352,7 +356,7 @@ def _take_along_dim_helper(self, indices, dim):
 def take_along_dim(input, indices, dim=None, *, out=None):
     if dim:
         self_broadcasted, indices_broadcasted, dim = _take_along_dim_helper(input, indices, dim)
-        return self_broadcasted.gather(dim, indices_broadcasted)
+        return gather(self_broadcasted, dim, indices_broadcasted)
     return input.view(-1).gather(0, indices.view(-1))
 
 # tensor_split
@@ -427,7 +431,7 @@ def where(condition, *args, out=None):
             input = mindspore.tensor(input, dtype=mindspore.float32)
         other = finfo(input.dtype).min
 
-    if use_pyboost():
+    if use_pyboost() and not ON_ORANGE_PI:
         output = mindspore.mint.where(condition, input, other)
     else:
         output = condition * input + (~condition) * other
diff --git a/mindnlp/core/ops/comparison.py b/mindnlp/core/ops/comparison.py
index 78cd343e0..18000f740 100644
--- a/mindnlp/core/ops/comparison.py
+++ b/mindnlp/core/ops/comparison.py
@@ -3,7 +3,7 @@
 import numpy as np
 import mindspore
 from mindspore import ops
-from ..configs import use_pyboost
+from ..configs import use_pyboost, ON_ORANGE_PI
 
 from ._inner import call_ms_func
 
@@ -64,7 +64,7 @@ def greater(input, other, *, out=None):
 # isclose
 has_isclose = hasattr(mindspore.mint, 'isclose')
 def isclose(input, other, rtol=1e-05, atol=1e-08, equal_nan=False):
-    if use_pyboost() and has_isclose:
+    if use_pyboost() and has_isclose and not ON_ORANGE_PI:
         return mindspore.mint.isclose(input, other, rtol, atol, equal_nan)
     return mindspore.tensor(np.isclose(input.numpy(), other.numpy(), rtol, atol, equal_nan))
 
@@ -174,7 +174,7 @@ def not_equal(input, other):
 # sort
 has_sort = hasattr(mindspore.mint, 'sort')
 def sort(input, *, dim=-1, descending=False, stable=False):
-    if use_pyboost() and has_sort:
+    if use_pyboost() and has_sort and not ON_ORANGE_PI:
         out = mindspore.mint.sort(input, dim=dim, descending=descending, stable=stable)
     else:
         out = ops.sort(input, dim, descending)
diff --git a/mindnlp/core/ops/pointwise.py b/mindnlp/core/ops/pointwise.py
index 8fe981939..98465b043 100644
--- a/mindnlp/core/ops/pointwise.py
+++ b/mindnlp/core/ops/pointwise.py
@@ -2,7 +2,7 @@
 import numpy as np
 import mindspore
 from mindspore import ops
-from ..configs import use_pyboost, ON_A1
+from ..configs import use_pyboost, ON_A1, ON_ORANGE_PI
 from ._inner import call_ms_func
 
 from mindnlp import core
@@ -582,7 +582,7 @@ def igammac(input, other):
 
 
 def mul(input, other, *, out=None):
-    if use_pyboost() and has_mul:
+    if use_pyboost() and has_mul and not ON_ORANGE_PI:
         out = mindspore.mint.mul(input, other)
     else:
         if input.dtype == mindspore.bool_:
diff --git a/mindnlp/utils/torch_proxy.py b/mindnlp/utils/torch_proxy.py
index ae1fccae8..1882dc0b2 100644
--- a/mindnlp/utils/torch_proxy.py
+++ b/mindnlp/utils/torch_proxy.py
@@ -62,6 +62,8 @@ def exec_module(self, module):
         class ProxyModule(type(module)):
             def __getattr__(_, name):
                 # 动态导入实际模块中的属性
+                if DEVICE_TARGET == 'Ascend':
+                    name = name.replace('cuda', 'npu')
                 try:
                     target_module = importlib.import_module(self.target_name)
                 except ImportError as e:
diff --git a/tests/run_test.py b/tests/run_test.py
index 6891cf5da..6bd7ff7d4 100644
--- a/tests/run_test.py
+++ b/tests/run_test.py
@@ -1,6 +1,5 @@
 import os
 import sys
-sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
 
 import pytest
 import mindspore