[AutoParallel] Set value phi yaml (#58893)

* set value to phi

paddle/fluid/eager/auto_code_generator/generator/eager_gen.py

@@ -826,7 +826,7 @@ def BackwardValidationCheck(self):
         max_grad_tensor_position = -1
         for _, (_, _, pos) in backward_grad_inputs_map.items():
             assert pos > max_fwd_input_position, AssertMessage(
-                pos, max_grad_tensor_position
+                pos, max_fwd_input_position
             )
             max_grad_tensor_position = max(max_grad_tensor_position, pos)
 

paddle/fluid/ir_adaptor/translator/op_translator.cc

@@ -2216,12 +2216,12 @@ struct SetValueWithTensorOpTranscriber : public SetValueOpTranscriber {
 struct SetValueGradOpTranscriber : public SetValueWithTensorOpTranscriber {
   pir::OpInfo LoopkUpOpInfo(pir::IrContext* ctx,
                             const OpDesc& op_desc) override {
-    std::string target_op_name = dialect::SetValueGradOp::name();
+    std::string target_op_name = dialect::SetValueWithTensorGradOp::name();
     const auto& op_info = ctx->GetRegisteredOpInfo(target_op_name);
     if (!op_info) {
       IR_THROW(
           "Op set_value_grad should have corresponding OpInfo "
-          "pd_op.set_value_grad");
+          "pd_op.set_value_with_tensor_grad");
     }
 
     return op_info;

paddle/fluid/pir/dialect/operator/ir/ops.yaml

@@ -140,30 +140,6 @@
     func : send_v2
     param : [x, ring_id, dynamic_shape, peer, use_calc_stream]
 
-- op : set_value
-  args : (Tensor x, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes, int64_t[] shape, Scalar[] values)
-  output : Tensor(out)
-  infer_meta:
-    func: SetValueInferMeta
-    param: [x]
-  kernel:
-    func: set_value
-    param: [x, starts, ends, steps, axes, decrease_axes, none_axes, shape, values]
-  inplace: (x -> out)
-  backward: set_value_grad
-
-- op : set_value_with_tensor
-  args : (Tensor x, Tensor values, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes)
-  output : Tensor(out)
-  infer_meta:
-    func: SetValueInferMeta
-    param: [x]
-  kernel:
-    func: set_value_with_tensor
-    param: [x, values, starts, ends, steps, axes, decrease_axes, none_axes]
-  inplace: (x -> out)
-  backward: set_value_grad
-
 - op : shadow_feed
   args : (Tensor x)
   output : Tensor(out)

paddle/fluid/pir/dialect/operator/ir/ops_backward.yaml

@@ -17,13 +17,3 @@
   kernel:
     func: fused_feedforward_grad
   optional: linear1_bias, linear2_bias, ln1_scale, ln1_bias, ln1_out, ln1_mean, ln1_variance, ln2_scale, ln2_bias, ln2_mean, ln2_variance, dropout2_out, ln1_scale_grad, ln1_bias_grad, ln2_scale_grad, ln2_bias_grad,  linear2_bias_grad
-
-- backward_op : set_value_grad
-  args : (Tensor out_grad, Tensor values, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes)
-  output : Tensor(x_grad), Tensor(values_grad)
-  infer_meta:
-    func: SetValueGradInferMeta
-    param: [out_grad, values]
-  kernel:
-    func: set_value_grad
-    param: [out_grad, starts, ends, steps, axes, decrease_axes, none_axes]

paddle/fluid/pybind/eager_method.cc

@@ -1306,7 +1306,7 @@ static PyObject* tensor__getitem_index_not_tensor(TensorObject* self,
   EAGER_TRY
   PyObject* _index = PyTuple_GET_ITEM(args, 0);
   VLOG(4) << "Call _getitem_index_not_tensor";
-  std::vector<int> slice_axes, slice_starts, slice_ends, slice_strides,
+  std::vector<int64_t> slice_axes, slice_starts, slice_ends, slice_strides,
       decrease_axis, none_axes, infer_flags;
   std::vector<int64_t> list_select_idxs;
   // if index is a list, list_select_flag will be true
@@ -1353,26 +1353,25 @@ static PyObject* tensor__getitem_index_not_tensor(TensorObject* self,
         break;
       }
     }
-    std::vector<int64_t> slice_axes_tmp(slice_axes.begin(), slice_axes.end());
-    std::vector<int64_t> infer_flags_tmp(infer_flags.begin(),
-                                         infer_flags.end());
-    std::vector<int64_t> decrease_axis_tmp(decrease_axis.begin(),
-                                           decrease_axis.end());
 
     if (op_type == "slice") {
       eager_gil_scoped_release guard;
       out = slice_ad_func(self->tensor,
-                          slice_axes_tmp,
+                          slice_axes,
                           slice_starts,
                           slice_ends,
-                          infer_flags_tmp,
-                          decrease_axis_tmp);
+                          infer_flags,
+                          decrease_axis);
     } else if (op_type == "strided_slice") {
       eager_gil_scoped_release guard;
-      out = strided_slice_ad_func(
-          self->tensor, slice_axes, slice_starts, slice_ends, slice_strides);
-      if (!decrease_axis_tmp.empty()) {
-        out = squeeze_ad_func(out, decrease_axis_tmp);
+      std::vector<int> slice_axes_tmp(slice_axes.begin(), slice_axes.end());
+      out = strided_slice_ad_func(self->tensor,
+                                  slice_axes_tmp,
+                                  slice_starts,
+                                  slice_ends,
+                                  slice_strides);
+      if (!decrease_axis.empty()) {
+        out = squeeze_ad_func(out, decrease_axis);
       }
     } else {
       PADDLE_THROW(platform::errors::InvalidArgument(
@@ -1607,7 +1606,7 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self,
   // TODO(liym27): Try not to call TensorToPyArray because it always
   // copys data to cpu place, which reduces performance.
   if (parse_index) {
-    std::vector<int> axes, starts, ends, steps, decrease_axes, none_axes,
+    std::vector<int64_t> axes, starts, ends, steps, decrease_axes, none_axes,
         infer_flags;
     std::vector<int64_t> list_select_idxs;
     // if index is a list, list_select_flag will be true
@@ -1624,13 +1623,6 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self,
                        &list_select_idxs,
                        &list_select_flag);
 
-    framework::AttributeMap attrs = {{"axes", axes},
-                                     {"starts", starts},
-                                     {"ends", ends},
-                                     {"steps", steps},
-                                     {"decrease_axes", decrease_axes},
-                                     {"none_axes", none_axes}};
-
     if (egr::Controller::Instance().HasGrad()) {
       PADDLE_ENFORCE_EQ(
           egr::EagerUtils::IsLeafTensor(self->tensor) &&
@@ -1643,6 +1635,8 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self,
     }
 
     paddle::Tensor value_tensor;
+    std::vector<phi::Scalar> values;
+    std::vector<int64_t> shape = std::vector<int64_t>{1};
 
     if (PyCheckTensor(value_obj)) {
       value_tensor = reinterpret_cast<TensorObject*>(value_obj)->tensor;
@@ -1706,25 +1700,20 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self,
           PyComplex_Check(value_obj)) {
         if (self->tensor.dtype() == phi::DataType::FLOAT32 ||
             self->tensor.dtype() == phi::DataType::FLOAT16) {
-          attrs["values"] = std::vector<paddle::experimental::Scalar>{
-              value_obj_tmp.cast<float>()};
+          values = std::vector<phi::Scalar>{value_obj_tmp.cast<float>()};
         } else if (self->tensor.dtype() == phi::DataType::FLOAT64) {
-          attrs["values"] = std::vector<paddle::experimental::Scalar>{
-              value_obj_tmp.cast<double>()};
+          values = std::vector<phi::Scalar>{value_obj_tmp.cast<double>()};
         } else if (self->tensor.dtype() == phi::DataType::INT32) {
-          attrs["values"] = std::vector<paddle::experimental::Scalar>{
-              value_obj_tmp.cast<int32_t>()};
+          values = std::vector<phi::Scalar>{value_obj_tmp.cast<int32_t>()};
         } else if (self->tensor.dtype() == phi::DataType::INT64) {
-          attrs["values"] = std::vector<paddle::experimental::Scalar>{
-              value_obj_tmp.cast<int64_t>()};
+          values = std::vector<phi::Scalar>{value_obj_tmp.cast<int64_t>()};
         } else if (self->tensor.dtype() == phi::DataType::BOOL) {
-          attrs["values"] = std::vector<paddle::experimental::Scalar>{
-              value_obj_tmp.cast<bool>()};
+          values = std::vector<phi::Scalar>{value_obj_tmp.cast<bool>()};
         } else if (self->tensor.dtype() == phi::DataType::COMPLEX64) {
-          attrs["values"] = std::vector<paddle::experimental::Scalar>{
+          values = std::vector<phi::Scalar>{
               value_obj_tmp.cast<std::complex<float>>()};
         } else if (self->tensor.dtype() == phi::DataType::COMPLEX128) {
-          attrs["values"] = std::vector<paddle::experimental::Scalar>{
+          values = std::vector<phi::Scalar>{
               value_obj_tmp.cast<std::complex<double>>()};
         } else {
           PADDLE_THROW(platform::errors::InvalidArgument(
@@ -1734,8 +1723,6 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self,
               "float16, "
               "please check the type of tensor."));
         }
-        attrs["shape"] = std::vector<int64_t>{1};
-
       } else {
         PADDLE_THROW(platform::errors::InvalidArgument(
             "Value type error. The assign value allows "
@@ -1748,25 +1735,46 @@ static PyObject* tensor_method__setitem_eager_tensor(TensorObject* self,
       // Release gil and do tracing
       py::gil_scoped_release release;
       // use inplace set_value_ operator
-      if (value_tensor.initialized() &&
-          (self->tensor.dtype() != value_tensor.dtype())) {
-        if (egr::Controller::Instance().GetAMPLevel() !=
-            paddle::imperative::AmpLevel::O0) {
-          paddle::small_vector<std::vector<paddle::Tensor>,
-                               egr::kSlotSmallVectorSize>
-              tmps = {{self->tensor}, {value_tensor}};
-          auto amp_dtype = egr::GetAmpDestDtype("set_value", tmps);
-          self->tensor = egr::EagerAmpAutoCast(
-              self->tensor.name(), self->tensor, amp_dtype, "set_value");
-          value_tensor = egr::EagerAmpAutoCast(
-              value_tensor.name(), value_tensor, amp_dtype, "set_value");
-        }
+      if (value_tensor.initialized()) {
         if (self->tensor.dtype() != value_tensor.dtype()) {
-          value_tensor = cast_ad_func(value_tensor, self->tensor.dtype());
+          if (egr::Controller::Instance().GetAMPLevel() !=
+              paddle::imperative::AmpLevel::O0) {
+            paddle::small_vector<std::vector<paddle::Tensor>,
+                                 egr::kSlotSmallVectorSize>
+                tmps = {{self->tensor}, {value_tensor}};
+            auto amp_dtype = egr::GetAmpDestDtype("set_value", tmps);
+            self->tensor = egr::EagerAmpAutoCast(
+                self->tensor.name(), self->tensor, amp_dtype, "set_value");
+            value_tensor = egr::EagerAmpAutoCast(
+                value_tensor.name(), value_tensor, amp_dtype, "set_value");
+          }
+          if (self->tensor.dtype() != value_tensor.dtype()) {
+            value_tensor = cast_ad_func(value_tensor, self->tensor.dtype());
+          }
+        }
+        const phi::distributed::ProcessMesh* mesh = nullptr;
+        if (InputsContainDistTensor(&mesh, self->tensor, value_tensor)) {
+          ConvertAllInputsToDistTensor(mesh, self->tensor, value_tensor);
         }
+        self->tensor = set_value_with_tensor__ad_func(self->tensor,
+                                                      value_tensor,
+                                                      starts,
+                                                      ends,
+                                                      steps,
+                                                      axes,
+                                                      decrease_axes,
+                                                      none_axes);
+      } else {
+        self->tensor = set_value__ad_func(self->tensor,
+                                          starts,
+                                          ends,
+                                          steps,
+                                          axes,
+                                          decrease_axes,
+                                          none_axes,
+                                          shape,
+                                          values);
       }
-      self->tensor = set_value__dygraph_function(
-          self->tensor, value_tensor, {}, {}, {}, attrs);
     }
     if (PyCheckTensor(value_obj)) {
       // pass the stop_gradient from value to tensor.

paddle/fluid/pybind/slice_utils.h

@@ -143,13 +143,13 @@ static int _PySlice_GetIndices(PySliceObject* r,
 
 static void ParseIndexingSlice(phi::DDim shape,
                                PyObject* _index,
-                               std::vector<int>* slice_axes,
-                               std::vector<int>* slice_starts,
-                               std::vector<int>* slice_ends,
-                               std::vector<int>* slice_strides,
-                               std::vector<int>* decrease_axis,
-                               std::vector<int>* none_axes,
-                               std::vector<int>* infer_flags,
+                               std::vector<int64_t>* slice_axes,
+                               std::vector<int64_t>* slice_starts,
+                               std::vector<int64_t>* slice_ends,
+                               std::vector<int64_t>* slice_strides,
+                               std::vector<int64_t>* decrease_axis,
+                               std::vector<int64_t>* none_axes,
+                               std::vector<int64_t>* infer_flags,
                                std::vector<int64_t>* list_select_idxs,
                                bool* list_select_flag) {
   // We allow indexing by Integers, Slices, Ellipsis, None, tuples of those

paddle/phi/api/yaml/legacy_backward.yaml

@@ -610,6 +610,28 @@
     func : rrelu_grad
     data_type : x
 
+- backward_op : set_value_grad
+  forward : set_value (Tensor x, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes, int64_t[] shape, Scalar[] values) -> Tensor(out)
+  args : (Tensor out_grad)
+  output : Tensor(x_grad)
+  infer_meta:
+    func: UnchangedInferMeta
+    param: [out_grad]
+  kernel:
+    func: assign
+    param: [out_grad]
+
+- backward_op : set_value_with_tensor_grad
+  forward: set_value_with_tensor (Tensor x, Tensor values, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes) -> Tensor(out)
+  args : (Tensor values,Tensor out_grad, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes)
+  output : Tensor(x_grad), Tensor(values_grad)
+  infer_meta:
+    func: SetValueGradInferMeta
+    param: [out_grad, values]
+  kernel:
+    func: set_value_grad
+    param: [out_grad, starts, ends, steps, axes, decrease_axes, none_axes]
+
 - backward_op : slice_double_grad
   forward : slice_grad (Tensor input, Tensor grad_out, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis) -> Tensor(grad_input)
   args : (Tensor grad_input_grad, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis)

paddle/phi/api/yaml/legacy_ops.yaml

@@ -961,6 +961,28 @@
   intermediate : noise
   backward : rrelu_grad
 
+- op : set_value
+  args : (Tensor x, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes, int64_t[] shape, Scalar[] values)
+  output : Tensor(out)
+  inplace: (x -> out)
+  infer_meta :
+    func : SetValueInferMeta
+    param : [x]
+  kernel :
+    func : set_value
+  backward: set_value_grad
+
+- op : set_value_with_tensor
+  args : (Tensor x, Tensor values, IntArray starts, IntArray ends, IntArray steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes)
+  output : Tensor(out)
+  inplace: (x -> out)
+  infer_meta:
+    func: SetValueInferMeta
+    param: [x]
+  kernel:
+    func: set_value_with_tensor
+  backward: set_value_with_tensor_grad
+
 - op : slice
   args : (Tensor input, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis)
   output : Tensor

python/paddle/base/variable_index.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import itertools
 import warnings
 from functools import reduce
 
@@ -875,6 +874,7 @@ def _setitem_static(x, indices, values):
     StartsTensorList = None
     EndsTensorList = None
     StepsTensorList = None
+    shape = None
 
     if paddle.utils._contain_var(starts):
         StartsTensorList = paddle.utils._convert_to_tensor_list(starts)
@@ -919,14 +919,29 @@ def _setitem_static(x, indices, values):
 
         # step3.1: Only basic indexing, use OP set_value to set value.
         if paddle.in_dynamic_mode():
-            return paddle._legacy_C_ops.set_value_(
-                x,
-                value_tensor,
-                StartsTensorList,
-                EndsTensorList,
-                StepsTensorList,
-                *itertools.chain.from_iterable(attrs.items()),
-            )
+            if value_tensor is None:
+                return paddle._C_ops.set_value_(
+                    x,
+                    starts,
+                    ends,
+                    steps,
+                    axes,
+                    decrease_axes,
+                    none_axes,
+                    shape,
+                    values,
+                )
+            else:
+                return paddle._C_ops.set_value_with_tensor_(
+                    x,
+                    value_tensor,
+                    starts,
+                    ends,
+                    steps,
+                    axes,
+                    decrease_axes,
+                    none_axes,
+                )
         else:
             helper = paddle.base.layer_helper.LayerHelper(
                 'set_value', **locals()

test/indexing/test_setitem.py

@@ -503,3 +503,7 @@ def test_indexing_is_multi_dim_list(self):
             res = self.exe.run(fetch_list=[y.name])
 
         np.testing.assert_allclose(res[0], np_data)
+
+
+if __name__ == '__main__':
+    unittest.main()