@@ -29,3 +58,3 @@
         super(Function, self).__init__()
+        self.ctx = FunctionCtx()
 
-    def apply(self, *args, **kwargs):

@@ -37,0 +70,4 @@
+    return input_ms, _origin_dtype
+
+# should use before cast_to_adapter_tensor
+def _recorver_dtype_on_ascend(output_ms, _origin_dtype):

@@ -492,2 +556,3 @@
         _bc_op = _get_cache_prim(ms.ops.Broadcast)(src, _group_name)
-    tensor.data = _bc_op((tensor,))[0]
+    if get_backend(group) == "hccl":
+        cast_tensor, _origin_dtype = _check_and_convert_dtype_on_ascend(tensor)

@@ -255,6 +255,13 @@ class _Optimizer:
         loss = None
         if closure is not None:
             loss = closure()
+        if grads is None:

@@ -258,0 +260,4 @@
+            for param_group in self.param_groups:
+                for param in param_group['param']:
+                    _grad = param.grad if param.grad is not None else ms.ops.zeros_like(param)
+                    grads.append(_grad)

@@ -42,0 +77,4 @@
+    def backward(ctx, *grad_outputs):
+        pass
+
+    def bprop(self, *args, **kwargs):

@@ -8,6 +8,8 @@ import mindspore as ms
 import numpy as np
 from mindspore import jit, grad
 
+user_stream1 = ms_torch.cuda.Stream()

@@ -36,0 +80,4 @@
+                                  "your custom autograd.Function to use it with backward "
+                                  "mode AD.")
+
+    def bprop(self, *args, **kwargs):

@@ -5,0 +25,4 @@
+        self.dirty_tensors = args
+
+    def mark_non_differentiable(self, *args):
+        self.non_differentiable = args