diff --git a/mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc b/mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc
index f01dd95f060be041f9bfc2ea46c376f7b8b12822..800c862c53f8b2738a31ad59fbbdd8d272d17625 100644
--- a/mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc
+++ b/mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc
@@ -100,6 +100,8 @@ void AddAscendBackendOptionalIRFusion(PassManager *ir_fusion_pm) {
   ir_fusion_pm->AddPass(std::make_shared<ClipByNormNoDivSquareSumFusion>());
   ir_fusion_pm->AddPass(std::make_shared<LambUpdateWithLRRuleFusion>());
   ir_fusion_pm->AddPass(std::make_shared<SoftmaxGradExtFusion>());
+  ir_fusion_pm->AddPass(std::make_shared<SoftmaxGradExtFusionV2>());
+  ir_fusion_pm->AddPass(std::make_shared<SoftmaxGradExtFusionV3>());
   ir_fusion_pm->AddPass(std::make_shared<ConfusionMulGradFusion>());
   ir_fusion_pm->AddPass(std::make_shared<ConfusionSoftmaxGradRule>());
   ir_fusion_pm->AddPass(std::make_shared<LambNextMVWithDecayRuleCond1>());
diff --git a/mindspore/ccsrc/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion.cc b/mindspore/ccsrc/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion.cc
index ccb0cbfcb8788afb90291f5691faf4bb68600bb5..429523bd8ba66ae730885838dc531f0134c62034 100644
--- a/mindspore/ccsrc/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion.cc
+++ b/mindspore/ccsrc/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion.cc
@@ -31,6 +31,24 @@ const BaseRef SoftmaxGradExtFusion::DefinePattern() const {
   return mul_grad;
 }
 
+const BaseRef SoftmaxGradExtFusionV2::DefinePattern() const {
+  VectorRef mul({prim::kPrimMul, input1_, input0_});
+  VectorRef sum({sum_var_, mul});
+  VectorRef sub({prim::kPrimSub, input0_, sum});
+  VectorRef mul1({prim::kPrimMul, input1_, sub});
+  VectorRef mul_grad({prim::kPrimMul, input2_, mul1});
+  return mul_grad;
+}
+
+const BaseRef SoftmaxGradExtFusionV3::DefinePattern() const {
+  VectorRef mul({prim::kPrimMul, input1_, input0_});
+  VectorRef sum({sum_var_, mul});
+  VectorRef sub({prim::kPrimSub, input0_, sum});
+  VectorRef mul1({prim::kPrimMul, input1_, sub});
+  VectorRef mul_grad({prim::kPrimMul, mul1, input2_});
+  return mul_grad;
+}
+
 const AnfNodePtr SoftmaxGradExtFusion::Process(const FuncGraphPtr &graph, const AnfNodePtr &node,
                                                const EquivPtr &equiv) const {
   MS_EXCEPTION_IF_NULL(graph);
@@ -46,7 +64,7 @@ const AnfNodePtr SoftmaxGradExtFusion::Process(const FuncGraphPtr &graph, const
   MS_EXCEPTION_IF_NULL(fusion_node);
   fusion_node->set_scope(node->scope());
   fusion_node->set_abstract(node->abstract());
-  AnfAlgo::CopyNodeAttr(kAttrKeepDims, sum, fusion_node);
+  AnfAlgo::CopyNodeAttr(kAttrKeepDims, "keepdims", sum, fusion_node);
   AnfAlgo::CopyNodeAttr(kAttrAxis, sum, fusion_node);
   return fusion_node;
 }
diff --git a/mindspore/ccsrc/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion.h b/mindspore/ccsrc/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion.h
index 70c5658e6086f42413ece583261f700ed7818aa7..59032e697335696203bb6dd6ac633a33c7eb3305 100644
--- a/mindspore/ccsrc/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion.h
+++ b/mindspore/ccsrc/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion.h
@@ -17,13 +17,15 @@
 #define MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_IR_FUSION_SOFTMAX_GRAD_EXT_FUSION_H_
 
 #include <memory>
+#include <string>
 #include "pre_activate/common/optimizer.h"
 
 namespace mindspore {
 namespace opt {
 class SoftmaxGradExtFusion : public PatternProcessPass {
  public:
-  explicit SoftmaxGradExtFusion(bool multigraph = true) : PatternProcessPass("softmax_grad_ext_fusion", multigraph) {
+  explicit SoftmaxGradExtFusion(const std::string &name = "softmax_grad_ext_fusion", bool multigraph = true)
+      : PatternProcessPass(name, multigraph) {
     input0_ = std::make_shared<Var>();
     input1_ = std::make_shared<Var>();
     input2_ = std::make_shared<Var>();
@@ -33,12 +35,28 @@ class SoftmaxGradExtFusion : public PatternProcessPass {
   const BaseRef DefinePattern() const override;
   const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
 
- private:
+ protected:
   VarPtr input0_;
   VarPtr input1_;
   VarPtr input2_;
   VarPtr sum_var_;
 };
+
+class SoftmaxGradExtFusionV2 : public SoftmaxGradExtFusion {
+ public:
+  explicit SoftmaxGradExtFusionV2(bool multigraph = true)
+      : SoftmaxGradExtFusion("softmax_grad_ext_fusion_v2", multigraph) {}
+  ~SoftmaxGradExtFusionV2() override = default;
+  const BaseRef DefinePattern() const override;
+};
+
+class SoftmaxGradExtFusionV3 : public SoftmaxGradExtFusion {
+ public:
+  explicit SoftmaxGradExtFusionV3(bool multigraph = true)
+      : SoftmaxGradExtFusion("softmax_grad_ext_fusion_v3", multigraph) {}
+  ~SoftmaxGradExtFusionV3() override = default;
+  const BaseRef DefinePattern() const override;
+};
 }  // namespace opt
 }  // namespace mindspore
 #endif  // MINDSPORE_CCSRC_PRE_ACTIVATE_ASCEND_IR_FUSION_SOFTMAX_GRAD_EXT_FUSION_H_
diff --git a/tests/ut/cpp/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion_test.cc b/tests/ut/cpp/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion_test.cc
index 25432336130956788222a1cfbde3107fee62627f..5f02f0e9c178a9573a4ce9bb493284f7d449d4d0 100644
--- a/tests/ut/cpp/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion_test.cc
+++ b/tests/ut/cpp/pre_activate/ascend/ir_fusion/softmax_grad_ext_fusion_test.cc
@@ -49,5 +49,47 @@ TEST_F(TestHWOptSoftmaxGradExtFusion, test_fusion) {
   FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_softmax_grad_ext_fusion", "after");
   EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
 }
+
+TEST_F(TestHWOptSoftmaxGradExtFusion, test_fusion_v2) {
+  FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_softmax_grad_ext_fusion_v2", "before");
+  EXPECT_NE(g, nullptr);
+  std::vector<int> shp{1, 1, 1, 1};
+  auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp);
+  AbstractBasePtrList args_spec_list;
+  for (size_t i = 0; i < 3; ++i) {
+    args_spec_list.push_back(x_abstract);
+  }
+  auto fg = GetKernelGraph(g, args_spec_list);
+
+  auto optimizer = std::make_shared<opt::GraphOptimizer>();
+  auto pm = std::make_shared<opt::PassManager>();
+  pm->AddPass(std::make_shared<opt::SoftmaxGradExtFusionV2>());
+  optimizer->AddPassManager(pm);
+  FuncGraphPtr new_graph = optimizer->Optimize(fg);
+
+  FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_softmax_grad_ext_fusion_v2", "after");
+  EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
+}
+
+TEST_F(TestHWOptSoftmaxGradExtFusion, test_fusion_v3) {
+  FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_softmax_grad_ext_fusion_v3", "before");
+  EXPECT_NE(g, nullptr);
+  std::vector<int> shp{1, 1, 1, 1};
+  auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp);
+  AbstractBasePtrList args_spec_list;
+  for (size_t i = 0; i < 3; ++i) {
+    args_spec_list.push_back(x_abstract);
+  }
+  auto fg = GetKernelGraph(g, args_spec_list);
+
+  auto optimizer = std::make_shared<opt::GraphOptimizer>();
+  auto pm = std::make_shared<opt::PassManager>();
+  pm->AddPass(std::make_shared<opt::SoftmaxGradExtFusionV3>());
+  optimizer->AddPassManager(pm);
+  FuncGraphPtr new_graph = optimizer->Optimize(fg);
+
+  FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_softmax_grad_ext_fusion_v3", "after");
+  EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
+}
 }  // namespace opt
 }  // namespace mindspore
diff --git a/tests/ut/cpp/python_input/gtest_input/pre_activate/softmax_grad_ext_fusion.py b/tests/ut/cpp/python_input/gtest_input/pre_activate/softmax_grad_ext_fusion.py
index fbcc3d7480d738da391e3b940113d4e57379b209..38b8be0493d90d4e15f817aa7ec68ba0985139f6 100644
--- a/tests/ut/cpp/python_input/gtest_input/pre_activate/softmax_grad_ext_fusion.py
+++ b/tests/ut/cpp/python_input/gtest_input/pre_activate/softmax_grad_ext_fusion.py
@@ -54,3 +54,43 @@ def test_softmax_grad_ext_fusion(tag):
         return MakeTuple(res)
 
     return fns[tag]
+
+
+def test_softmax_grad_ext_fusion_v2(tag):
+    fns = FnDict()
+
+    @fns
+    def before(input0, input1, input2):
+        mul = Mul(input1, input0)
+        reduce_sum = ReduceSum(mul, axes)
+        sub = Sub(input0, reduce_sum)
+        mul1 = Mul(input1, sub)
+        mul_grad = Mul(input2, mul1)
+        return mul_grad
+
+    @fns
+    def after(input0, input1, input2):
+        res = SoftmaxGradExt(input0, input1, input2)
+        return MakeTuple(res)
+
+    return fns[tag]
+
+
+def test_softmax_grad_ext_fusion_v3(tag):
+    fns = FnDict()
+
+    @fns
+    def before(input0, input1, input2):
+        mul = Mul(input1, input0)
+        reduce_sum = ReduceSum(mul, axes)
+        sub = Sub(input0, reduce_sum)
+        mul1 = Mul(input1, sub)
+        mul_grad = Mul(mul1, input2)
+        return mul_grad
+
+    @fns
+    def after(input0, input1, input2):
+        res = SoftmaxGradExt(input0, input1, input2)
+        return MakeTuple(res)
+
+    return fns[tag]