[Dy2static] support set_value_op in static mode by _jst.Ld() (#56028)

* Fix and add unittest

* don't introduce assign when already in global block.

* fix more

* fix bugs

* fix bugs

* fix ci

* fix bfgs

* make function local

python/paddle/fluid/framework.py

@@ -3573,6 +3573,10 @@ def _stride_in_no_check_dy2st_diff():
 
 
 def check_if_to_static_diff_with_dygraph(op_type, inplace_map, outputs):
+    if (
+        op_type == "while"
+    ):  # dont' need check while, while is only a wrapper of inner ops, we will stuck in inner op.
+        return
     if outputs is not None:
         for k, v in outputs.items():
             if isinstance(v, Variable):

python/paddle/fluid/layer_helper_base.py

@@ -466,6 +466,39 @@ class LayerHelperBase:
             stop_gradient=stop_gradient,
         )
 
+    def _create_global_variable_for_type_inference(
+        self, dtype, stop_gradient=False, shape=None
+    ):
+        """Create a global variable that should be type inferred layer.
+
+        Note:
+            The default type will be set to LOD_TENSOR. However, when
+            the var is used as operator output, its type will be updated
+            based on operator's `VarTypeInference` implementation in
+            infer_var_type.
+        """
+        # set global dtype
+        if not dtype:
+            dtype = self.__dtype
+        output = self.main_program.global_block().create_var(
+            name=unique_name.generate_with_ignorable_key(
+                ".".join([self.name, 'tmp'])
+            ),
+            dtype=dtype,
+            shape=shape,
+            type=core.VarDesc.VarType.LOD_TENSOR,
+            persistable=False,
+            stop_gradient=stop_gradient,
+        )
+        saved_block_id = self.main_program.current_block_idx
+        self.main_program.current_block_idx = 0
+        paddle.tensor.creation.fill_constant(
+            output.shape, dtype, 0.0, force_cpu=False, out=output
+        )
+        output.stop_gradient = stop_gradient
+        self.main_program.current_block_idx = saved_block_id
+        return output
+
     def create_sparse_variable_for_type_inference(
         self, dtype, stop_gradient=False, shape=None
     ):

python/paddle/fluid/variable_index.py

@@ -566,6 +566,12 @@ def _setitem_impl_(var, item, value):
         output = var
     else:
         helper = paddle.fluid.layer_helper.LayerHelper('set_value', **locals())
+        if helper.main_program.current_block_idx != 0:
+            # not in global block, we should create a global variable.
+            output = helper._create_global_variable_for_type_inference(
+                dtype=var.dtype
+            )
+        else:
             output = helper.create_variable_for_type_inference(dtype=var.dtype)
 
     cur_block = default_main_program().current_block()
@@ -909,7 +915,15 @@ def _setitem_static(x, indices, values):
             helper = paddle.fluid.layer_helper.LayerHelper(
                 'set_value', **locals()
             )
-            output = helper.create_variable_for_type_inference(dtype=x.dtype)
+            if helper.main_program.current_block_idx != 0:
+                # not in global block, we should create a global variable.
+                output = helper._create_global_variable_for_type_inference(
+                    dtype=x.dtype
+                )
+            else:
+                output = helper.create_variable_for_type_inference(
+                    dtype=x.dtype
+                )
         cur_block = default_main_program().current_block()
         cur_block.append_op(
             type="set_value",
@@ -975,7 +989,15 @@ def _setitem_static(x, indices, values):
             helper = paddle.fluid.layer_helper.LayerHelper(
                 'set_value', **locals()
             )
-            output = helper.create_variable_for_type_inference(dtype=x.dtype)
+            if helper.main_program.current_block_idx != 0:
+                # not in global block, we should create a global variable.
+                output = helper._create_global_variable_for_type_inference(
+                    dtype=x.dtype
+                )
+            else:
+                output = helper.create_variable_for_type_inference(
+                    dtype=x.dtype
+                )
         cur_block = default_main_program().current_block()
         cur_block.append_op(
             type="set_value",

python/paddle/jit/dy2static/basic_api_transformer.py

@@ -152,11 +152,14 @@ class NameloadJstTransformer(BaseTransformer):
         Can't convert name of function call, bacause this will affect CallTransformer.
         """
         node.args = [self.generic_visit(arg) for arg in node.args]
+        node.func = self.generic_visit(node.func)
         return node
 
     def visit_Attribute(self, node):
         assert isinstance(node, gast.Attribute)
         assert isinstance(node.attr, str)
+        if utils.ast_to_source_code(node).startswith("_jst."):  # skip _jst.xxx
+            return node
         self.generic_visit(node)
         if isinstance(node.ctx, gast.Load):
             node = self._surround_with_ld(node)

python/paddle/jit/dy2static/program_translator.py

@@ -1296,9 +1296,16 @@ class ParametersMap:
         params = self.params_dict.get(self._program_hash(program))
         if params is None:
             return None
-        if id in params.keys():
-            return params[id]
+        if id not in params:
             return None
+        root_var = params[id]
+        saved = []
+        while root_var.desc.id() in params.keys():
+            saved.append(root_var)
+            root_var = params[root_var.desc.id()]
+        for var in saved:
+            params[var.desc.id()] = root_var
+        return root_var
 
     def _program_hash(self, program):
         """

python/paddle/static/nn/control_flow.py

@@ -589,6 +589,7 @@ def assign_skip_lod_tensor_array(input, output):
         # input is not generated in While sub block and modified by in-place and only
         # belong to inplace ops in constructing program process, because in-place pass
         # is only available in Graph level.
+        with paddle.fluid.framework._stride_in_no_check_dy2st_diff():
             paddle.assign(input, output)
 
 

test/dygraph_to_static/test_jit_setitem.py

@@ -179,22 +179,56 @@ class TestCase11(TestSetItemBase):
 
 
 class TestCase12(TestSetItemBase):
-    # Test combind-indexing
+    # Test gradient of value tensor
     def init_func(self):
-        def foo(x, value):
-            y = x + 1
-            y[[0, 1], 1, :2] = value
-            return y
+        def foo():
+            res = paddle.zeros([4, 3, 2])
+            b = paddle.zeros([4, 3, 2])
+            v = paddle.to_tensor(1.0)
+            for i in range(paddle.shape(b)[0]):
+                res[i] = v
+            return res
 
         return foo
 
     def run_dygraph(self, func):
-        x = self.init_data()
-        value = paddle.ones((32,))
-        value.stop_gradient = False
-        y = func(x, value)
-        x_grad, value_grad = paddle.grad(y, [x, value])
-        return y, x_grad, value_grad
+        y = func()
+        return (y,)
+
+
+class TestCase13(TestSetItemBase):
+    # Test gradient of value tensor
+    def init_func(self):
+        def foo():
+            res = paddle.zeros([4, 3, 2])
+            v = paddle.to_tensor(1.0)
+            for i in range(4):
+                res[i] = v
+            return res
+
+        return foo
+
+    def run_dygraph(self, func):
+        y = func()
+        return (y,)
+
+
+class TestCase14(TestSetItemBase):
+    # Test gradient of value tensor
+    def init_func(self):
+        def foo():
+            data = np.arange(8).reshape((2, 4)).astype('float32')
+            x = paddle.to_tensor(data)
+            x[:, 1:] = x[:, :-1].clone()
+            x[:, 0] = 1
+            res = x.flatten()
+            return res
+
+        return foo
+
+    def run_dygraph(self, func):
+        y = func()
+        return (y,)
 
 
 if __name__ == '__main__':