PaddlePaddle
diff --git a/‎paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/same_operands_result.cc
Copy file name to clipboard
+2Lines changed: 2 additions & 0 deletions b/‎paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/same_operands_result.cc
Copy file name to clipboard
+2Lines changed: 2 additions & 0 deletions
diff --git a/‎paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/same_operands_result.h
Copy file name to clipboard
+2Lines changed: 2 additions & 0 deletions b/‎paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/same_operands_result.h
Copy file name to clipboard
+2Lines changed: 2 additions & 0 deletions
diff --git a/‎paddle/fluid/pybind/eager_method.cc
Copy file name to clipboard
+4-7Lines changed: 4 additions & 7 deletions b/‎paddle/fluid/pybind/eager_method.cc
Copy file name to clipboard
+4-7Lines changed: 4 additions & 7 deletions
diff --git a/‎paddle/fluid/pybind/slice_utils.h
Copy file name to clipboard
+19Lines changed: 19 additions & 0 deletions b/‎paddle/fluid/pybind/slice_utils.h
Copy file name to clipboard
+19Lines changed: 19 additions & 0 deletions
diff --git a/‎paddle/phi/infermeta/binary.cc
Copy file name to clipboard
+8Lines changed: 8 additions & 0 deletions b/‎paddle/phi/infermeta/binary.cc
Copy file name to clipboard
+8Lines changed: 8 additions & 0 deletions
diff --git a/‎paddle/phi/infermeta/binary.h
Copy file name to clipboard
+5Lines changed: 5 additions & 0 deletions b/‎paddle/phi/infermeta/binary.h
Copy file name to clipboard
+5Lines changed: 5 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/cpu/masked_fill_grad_kernel.cc
Copy file name to clipboard
+115Lines changed: 115 additions & 0 deletions b/‎paddle/phi/kernels/cpu/masked_fill_grad_kernel.cc
Copy file name to clipboard
+115Lines changed: 115 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/cpu/masked_fill_kernel.cc
Copy file name to clipboard
+93Lines changed: 93 additions & 0 deletions b/‎paddle/phi/kernels/cpu/masked_fill_kernel.cc
Copy file name to clipboard
+93Lines changed: 93 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/funcs/masked_fill_utils.h
Copy file name to clipboard
+37Lines changed: 37 additions & 0 deletions b/‎paddle/phi/kernels/funcs/masked_fill_utils.h
Copy file name to clipboard
+37Lines changed: 37 additions & 0 deletions
@@ -208,6 +208,8 @@ OP_SAME_OPERANDS_AND_RESULT(Renorm_)
 OP_SAME_OPERANDS_AND_RESULT(TanhShrink)
 OP_SAME_OPERANDS_AND_RESULT(YoloBoxHead)
 OP_SAME_OPERANDS_AND_RESULT(StandardGamma)
+OP_SAME_OPERANDS_AND_RESULT(MaskedFill)
+OP_SAME_OPERANDS_AND_RESULT(MaskedFill_)
 
 bool ScaleOpInferSymbolicShape(pir::Operation *op,
                                pir::InferSymbolicShapeContext *infer_context) {
 
@@ -205,6 +205,8 @@ OP_DECLARE_INFER_SYMBOLIC_SHAPE(Renorm_)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(TanhShrink)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(YoloBoxHead)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(StandardGamma)
+OP_DECLARE_INFER_SYMBOLIC_SHAPE(MaskedFill)
+OP_DECLARE_INFER_SYMBOLIC_SHAPE(MaskedFill_)
 
 }  // namespace paddle::dialect
 
 
@@ -1928,15 +1928,12 @@ static PyObject* tensor__setitem_dygraph(TensorObject* self,
         ConvertAllInputsToDistTensor(
             mesh, self->tensor, transed_sub_tensor, value_tensor);
       }
+      paddle::Tensor mask_tensor;
       if (transed_index.size() == 1 &&
-          transed_index[0].dtype() == phi::DataType::BOOL &&
-          transed_index[0].shape().size() == self->tensor.shape().size() &&
-          value_tensor.numel() == 1) {
-        value_tensor = expand_ad_func(value_tensor, self->tensor.shape());
+          MaskedFillDispatching(
+              transed_sub_tensor, value_tensor, transed_index, &mask_tensor)) {
         transed_sub_tensor =
-            where__ad_func(logical_not_ad_func(transed_index[0]),
-                           transed_sub_tensor,
-                           value_tensor);
+            masked_fill__ad_func(transed_sub_tensor, mask_tensor, value_tensor);
       } else {
         transed_sub_tensor =
             index_put__ad_func(transed_sub_tensor, transed_index, value_tensor);
 
@@ -409,6 +409,25 @@ static paddle::Tensor getTensorWithBasicIndexing(
   return out;
 }
 
+inline static bool MaskedFillDispatching(
+    const paddle::Tensor& tensor,
+    const paddle::Tensor& value,
+    const std::vector<paddle::Tensor>& indices,
+    paddle::Tensor* mask_tensor) {
+  if (value.numel() != 1) return false;
+  int64_t num_ind = 0;
+  if ((indices)[0].dtype() != phi::DataType::BOOL) {
+    return false;
+  } else {
+    num_ind += (indices)[0].shape().size();
+  }
+  *mask_tensor = (indices)[0];
+  for (size_t i = num_ind; i < tensor.shape().size(); i++) {
+    *mask_tensor = unsqueeze_ad_func(*mask_tensor, {-1});
+  }
+  return true;
+}
+
 static paddle::Tensor dealWithAdvancedIndex(
     const paddle::Tensor& tensor,
     std::vector<int>* advanced_index_dim,
 
@@ -2902,6 +2902,14 @@ void MaskedSelectInferMeta(const MetaTensor& x,
   out->set_dtype(x.dtype());
 }
 
+void MaskedFillInferMeta(const MetaTensor& x,
+                         const MetaTensor& mask,
+                         const MetaTensor& value,
+                         MetaTensor* out) {
+  out->set_dims(x.dims());  // can not infer
+  out->set_dtype(x.dtype());
+}
+
 void MatmulInferMeta(const MetaTensor& x,
                      const MetaTensor& y,
                      bool trans_x,
 
@@ -529,6 +529,11 @@ void MaskedSelectInferMeta(const MetaTensor& x,
                            const MetaTensor& mask,
                            MetaTensor* out);
 
+void MaskedFillInferMeta(const MetaTensor& x,
+                         const MetaTensor& mask,
+                         const MetaTensor& value,
+                         MetaTensor* out);
+
 void MatmulInferMeta(const MetaTensor& x,
                      const MetaTensor& y,
                      bool trans_x,
 
@@ -0,0 +1,115 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/phi/kernels/masked_fill_grad_kernel.h"
+
+#include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/phi/common/amp_type_traits.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/empty_kernel.h"
+#include "paddle/phi/kernels/expand_grad_kernel.h"
+#include "paddle/phi/kernels/expand_kernel.h"
+#include "paddle/phi/kernels/funcs/common_infer_shape_functions.h"
+#include "paddle/phi/kernels/funcs/common_shape.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void MaskedFillGradKernel(const Context& dev_ctx,
+                          const DenseTensor& x,
+                          const DenseTensor& mask,
+                          const DenseTensor& value UNUSED,
+                          const DenseTensor& out_grad,
+                          DenseTensor* x_grad,
+                          DenseTensor* v_grad) {
+  auto x_grad_dims = x_grad->dims();
+  auto mask_dims = mask.dims();
+  bool expand_x = false;
+  auto expanded_size =
+      common::vectorize(funcs::BroadcastTwoDims(x_grad_dims, mask_dims, -1));
+
+  DenseTensor mask_expand;
+  DenseTensor x_grad_expand;
+
+  auto expanded_dims = common::make_ddim(expanded_size);
+
+  if (mask.dims() != expanded_dims) {
+    ExpandKernel<bool, Context>(
+        dev_ctx, mask, IntArray(expanded_size), &mask_expand);
+  } else {
+    mask_expand = mask;
+  }
+
+  if (x_grad->dims() != expanded_dims) {
+    x_grad_expand = Empty<T, Context>(dev_ctx, IntArray(expanded_size));
+    expand_x = true;
+  } else {
+    x_grad_expand = *x_grad;
+  }
+
+  auto* mask_data = mask_expand.data<bool>();
+  auto* dout = out_grad.data<T>();
+  auto numel = mask_expand.numel();
+
+  if (x_grad != nullptr) {
+    dev_ctx.template Alloc<T>(x_grad);
+    auto mask_size = mask_expand.numel();
+    if (mask_size <= 0) return;
+
+    DenseTensor* x_grad_tmp = x_grad;
+    if (expand_x) {
+      x_grad_tmp = &x_grad_expand;
+    }
+    auto* dx = x_grad_tmp->data<T>();
+    for (int i = 0; i < numel; i++) {
+      dx[i] = mask_data[i] ? T{} : dout[i];
+    }
+
+    if (expand_x) {
+      ExpandGradKernel<T, Context>(
+          dev_ctx, x, x_grad_expand, IntArray(expanded_size), x_grad);
+    }
+  }
+
+  if (v_grad != nullptr) {
+    auto* dv = dev_ctx.template Alloc<int64_t>(v_grad);
+    dv[0] = 0;
+    for (int i = 0; i < numel; i++) {
+      if (mask_data[i]) {
+        dv[0]++;
+      }
+    }
+  }
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(masked_fill_grad,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::MaskedFillGradKernel,
+                   bool,
+                   float,
+                   double,
+                   int,
+                   int8_t,
+                   int64_t,
+                   int16_t,
+                   uint8_t,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16,
+                   phi::dtype::complex<float>,
+                   phi::dtype::complex<double>) {
+  kernel->InputAt(1).SetDataType(phi::DataType::BOOL);
+}
@@ -0,0 +1,93 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "paddle/phi/kernels/masked_fill_kernel.h"
+
+#include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/phi/common/amp_type_traits.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/expand_kernel.h"
+#include "paddle/phi/kernels/funcs/common_infer_shape_functions.h"
+#include "paddle/phi/kernels/funcs/common_shape.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void MaskedFillKernel(const Context& dev_ctx,
+                      const DenseTensor& x,
+                      const DenseTensor& mask,
+                      const DenseTensor& value,
+                      DenseTensor* out) {
+  auto x_dims = x.dims();
+  auto mask_dims = mask.dims();
+
+  auto expanded_size =
+      common::vectorize(phi::funcs::BroadcastTwoDims(x_dims, mask_dims, -1));
+
+  DenseTensor mask_expand;
+  DenseTensor x_expand;
+  auto value_data = value.data<T>();
+
+  DDim expand_dims = common::make_ddim(expanded_size);
+  if (mask.dims() != expand_dims) {
+    ExpandKernel<bool, Context>(
+        dev_ctx, mask, IntArray(expanded_size), &mask_expand);
+  } else {
+    mask_expand = mask;
+  }
+
+  if (x.dims() != expand_dims) {
+    ExpandKernel<T, Context>(dev_ctx, x, IntArray(expanded_size), &x_expand);
+  } else {
+    x_expand = x;
+  }
+
+  auto input_data = x_expand.data<T>();
+  auto mask_data = mask_expand.data<bool>();
+
+  auto x_size = x_expand.numel();
+
+  out->Resize(expand_dims);
+
+  auto out_data = dev_ctx.template HostAlloc<T>(out);
+
+  for (int i = 0; i < x_size; i++) {
+    if (mask_data[i]) {
+      out_data[i] = value_data[0];
+    } else {
+      out_data[i] = input_data[i];
+    }
+  }
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(masked_fill,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::MaskedFillKernel,
+                   bool,
+                   float,
+                   double,
+                   int,
+                   int8_t,
+                   int64_t,
+                   int16_t,
+                   uint8_t,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16,
+                   phi::dtype::complex<float>,
+                   phi::dtype::complex<double>) {
+  kernel->InputAt(1).SetDataType(phi::DataType::BOOL);
+}
@@ -0,0 +1,37 @@
+// Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+namespace phi {
+namespace funcs {
+inline bool CanDispatchMaskFillShortcut(const DDim& input_dim,
+                                        const DDim& mask_dim) {
+  int64_t broadcast_dim = -1;
+  for (size_t i = 0; i < input_dim.size(); i++) {
+    if (input_dim[i] != mask_dim[i]) {
+      broadcast_dim = i;
+      break;
+    }
+  }
+
+  for (size_t i = broadcast_dim; i < mask_dim.size(); i++) {
+    if (mask_dim[i] != 1) {
+      return false;
+    }
+  }
+  return true;
+}
+}  // namespace funcs
+}  // namespace phi