[API-Compat] Add paddle.compat.min/max and new PHI kernel (min/max_with_index)

paddle/fluid/pir/dialect/op_generator/op_build_gen.py

@@ -135,6 +135,7 @@
     'KthvalueInferMeta',
     'MaxPoolWithIndexInferMeta',
     'MaxPoolV2InferMeta',
+    'MinMaxWithIndexInferMeta',
     'MultinomialInferMeta',
     'OverlapAddInferMeta',
     'PadInferMeta',

paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/unary_infer_sym.cc

@@ -315,19 +315,37 @@ bool AnyOpInferSymbolicShape(pir::Operation *op,
                                  axis.size() == 0 /*reduce_all*/);
 }
 
-bool ArgmaxOpInferSymbolicShape(pir::Operation *op,
-                                pir::InferSymbolicShapeContext *infer_context) {
+bool MinMaxOpInferSymbolicShape(pir::Operation *op,
+                                pir::InferSymbolicShapeContext *infer_context,
+                                bool output_val_and_ind = false) {
   bool flatten = GetBoolAttr(op, "flatten");
-  bool keepdims = GetBoolAttr(op, "keepdims");
+  bool keepdims = false;
+  int axis = 0;
+
+  if (output_val_and_ind) {
+    keepdims = GetBoolAttr(op, "keepdim");
 
+    PADDLE_ENFORCE_NE(
+        op->attributes().find("dim"),
+        op->attributes().end(),
+        common::errors::InvalidArgument(
+            "'dim' Attribute is expected for Min/MaxWithIndexOp. "));
+    axis = op->attributes()
+               .at("dim")
+               .dyn_cast<paddle::dialect::ScalarAttribute>()
+               .data()
+               .to<int64_t>();
+  } else {
+    keepdims = GetBoolAttr(op, "keepdims");
+    const auto &axis_shape_or_data =
+        infer_context->GetShapeOrDataForValue(op->operand_source(1));
+    axis = static_cast<int>(
+        axis_shape_or_data.data().value().at(0).Get<int64_t>());
+  }
   const auto &input_sym_shape =
       infer_context->GetShapeOrDataForValue(op->operand_source(0)).shape();
-  int rank = input_sym_shape.size();
 
-  const auto &axis_shape_or_data =
-      infer_context->GetShapeOrDataForValue(op->operand_source(1));
-  int axis =
-      static_cast<int>(axis_shape_or_data.data().value().at(0).Get<int64_t>());
+  int rank = input_sym_shape.size();
   if (axis < 0) axis += rank;
 
   const auto &out_sym_shape = [&] {
@@ -357,14 +375,31 @@ bool ArgmaxOpInferSymbolicShape(pir::Operation *op,
       symbol::TensorShapeOrDataDimExprs(out_sym_shape)};
 
   infer_context->SetShapeOrDataForValue(op->result(0), shape_data);
+  if (output_val_and_ind)
+    infer_context->SetShapeOrDataForValue(op->result(1), shape_data);
   return true;
 }
 
+#define DEFINE_MINMAX_OP_INFER_FUNC(OpName, output_val_and_ind)               \
+  bool OpName##OpInferSymbolicShape(                                          \
+      pir::Operation *op, pir::InferSymbolicShapeContext *infer_context) {    \
+    return MinMaxOpInferSymbolicShape(op, infer_context, output_val_and_ind); \
+  }
+
+DEFINE_MINMAX_OP_INFER_FUNC(Argmax, false)
+DEFINE_MINMAX_OP_INFER_FUNC(MaxWithIndex, true)
+#undef DEFINE_MINMAX_OP_INFER_FUNC
+
 bool ArgminOpInferSymbolicShape(pir::Operation *op,
                                 pir::InferSymbolicShapeContext *infer_context) {
   return ArgmaxOpInferSymbolicShape(op, infer_context);
 }
 
+bool MinWithIndexOpInferSymbolicShape(
+    pir::Operation *op, pir::InferSymbolicShapeContext *infer_context) {
+  return MaxWithIndexOpInferSymbolicShape(op, infer_context);
+}
+
 bool AsComplexOpInferSymbolicShape(
     pir::Operation *op, pir::InferSymbolicShapeContext *infer_context) {
   pir::Value operand_source = op->operand_source(0);

paddle/fluid/pir/dialect/operator/interface/infer_symbolic_shape/unary_infer_sym.h

@@ -93,8 +93,10 @@ OP_DECLARE_INFER_SYMBOLIC_SHAPE(Lu)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(Lu_)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(Mode)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(Max)
+OP_DECLARE_INFER_SYMBOLIC_SHAPE(MaxWithIndex)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(Maxout)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(Min)
+OP_DECLARE_INFER_SYMBOLIC_SHAPE(MinWithIndex)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(Mean)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(MeanAll)
 OP_DECLARE_INFER_SYMBOLIC_SHAPE(MatrixPower)

paddle/phi/infermeta/unary.cc

@@ -366,6 +366,90 @@ void ArgMinMaxInferMeta(const MetaTensor& x,
   }
 }
 
+void MinMaxWithIndexInferMeta(const MetaTensor& x,
+                              const Scalar& axis,
+                              bool keepdims,
+                              bool flatten,
+                              MetaTensor* val_out,
+                              MetaTensor* ind_out,
+                              MetaConfig config) {
+  DataType val_dtype = x.dtype();
+
+  if (!config.is_runtime && axis.FromTensor()) {
+    std::vector<int64_t> vec;
+    if (flatten) {
+      if (keepdims) {  // NOLINT
+        vec = std::vector<int64_t>(x.dims().size(), -1);
+      } else {
+        vec = {};
+      }
+    } else {
+      if (keepdims) {
+        vec = std::vector<int64_t>(x.dims().size(), -1);
+      } else {
+        vec = std::vector<int64_t>(x.dims().size() - 1, -1);
+      }
+    }
+    val_out->set_dims(common::make_ddim(vec));
+    val_out->set_dtype(val_dtype);
+    ind_out->set_dims(common::make_ddim(vec));
+    ind_out->set_dtype(DataType::INT64);
+    return;
+  }
+  auto int_axis = axis.to<int64_t>();
+  const auto& x_dims = x.dims();
+
+  auto x_rank = x.dims().size();
+  if (x_rank > 0) {
+    PADDLE_ENFORCE_GE(int_axis,
+                      -x_rank,
+                      common::errors::InvalidArgument(
+                          "'axis'(%d) must be greater than or equal to"
+                          " -Rank(X)(%d).",
+                          int_axis,
+                          -x_rank));
+    PADDLE_ENFORCE_LT(
+        int_axis,
+        x_rank,
+        common::errors::InvalidArgument(
+            "'axis'(%d) must be less than Rank(X)(%d) of Input(X).",
+            int_axis,
+            x_rank));
+  } else {
+    // 0-dim tensor
+    PADDLE_ENFORCE_EQ(int_axis == 0 || int_axis == -1,
+                      true,
+                      common::errors::InvalidArgument(
+                          "'axis'(%d) must be 0 or -1 if input tensor is "
+                          "0-dim.",
+                          int_axis));
+  }
+
+  if (int_axis < 0) int_axis += x_rank;
+
+  std::vector<int64_t> vec;
+  if (flatten) {
+    if (keepdims) {  // NOLINT
+      vec = std::vector<int64_t>(x.dims().size(), 1);
+    } else {
+      vec = {};
+    }
+  } else {
+    for (int64_t i = 0; i < int_axis; i++)
+      vec.emplace_back(x_dims[static_cast<int>(i)]);
+    if (keepdims) {
+      vec.emplace_back(static_cast<int64_t>(1));
+    }
+    for (int64_t i = int_axis + 1; i < x_rank; i++)
+      vec.emplace_back(x_dims[static_cast<int>(i)]);
+  }
+
+  val_out->set_dims(common::make_ddim(vec));
+  val_out->set_dtype(val_dtype);
+  ind_out->set_dims(common::make_ddim(vec));
+  ind_out->set_dtype(DataType::INT64);
+}
+
 void ArgsortInferMeta(const MetaTensor& input,
                       int axis,
                       bool descending,

paddle/phi/infermeta/unary.h

@@ -66,6 +66,14 @@ void ArgMinMaxInferMeta(const MetaTensor& x,
                         MetaTensor* out,
                         MetaConfig config = MetaConfig());
 
+void MinMaxWithIndexInferMeta(const MetaTensor& x,
+                              const Scalar& axis,
+                              bool keepdims,
+                              bool flatten,
+                              MetaTensor* val_out,
+                              MetaTensor* ind_out,
+                              MetaConfig config = MetaConfig());
+
 void ArgsortInferMeta(const MetaTensor& input,
                       int axis,
                       bool descending,

paddle/phi/kernels/gpu/min_max_with_index_grad_kernel.cu

@@ -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
+//
+//     http://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/backends/gpu/gpu_context.h"
+#include "paddle/phi/common/place.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/core/utils/data_type.h"
+#include "paddle/phi/kernels/funcs/gather_scatter_functor.h"
+#include "paddle/phi/kernels/funcs/math_function.h"
+
+namespace phi {
+
+template <typename T>
+using EnableIfInteger =
+    typename std::enable_if<std::is_integral<T>::value, int>::type;
+
+template <typename T>
+using EnableIfNonInteger =
+    typename std::enable_if<!std::is_integral<T>::value, int>::type;
+
+// Here if keepdim=True, this will fallback to a simplified version of
+// take_along_axis. However, if keepdim=False (by default), indices will
+// not have equal rank will the input values (and values_grad), therefore
+// needs an unsqueeze operation by shallow copying indices and Resize
+#define DEFINE_WITH_INDEX_GRAD_KERNEL(OpType)                                \
+  template <typename T, typename Context, EnableIfNonInteger<T> = 0>         \
+  void OpType##WithIndexGradKernel(const Context& dev_ctx,                   \
+                                   const DenseTensor& x,                     \
+                                   const DenseTensor& values,                \
+                                   const DenseTensor& indices,               \
+                                   const DenseTensor& values_grad,           \
+                                   const Scalar& dim,                        \
+                                   bool keepdim,                             \
+                                   DenseTensor* x_grad) {                    \
+    x_grad->Resize(x.dims());                                                \
+    dev_ctx.template Alloc<T>(x_grad);                                       \
+    if (x_grad->numel() == 0) {                                              \
+      return;                                                                \
+    }                                                                        \
+    int64_t dim_val = dim.to<int64_t>();                                     \
+    if (dim_val < 0) {                                                       \
+      dim_val += x.dims().size();                                            \
+    }                                                                        \
+    DenseTensor shallow_copied_inds(indices);                                \
+    if (!keepdim) {                                                          \
+      auto indices_dim = x.dims();                                           \
+      indices_dim[dim_val] = 1;                                              \
+      shallow_copied_inds.Resize(indices_dim);                               \
+    }                                                                        \
+    phi::funcs::SetConstant<Context, T> functor;                             \
+    functor(dev_ctx, x_grad, static_cast<T>(0));                             \
+    phi::funcs::gpu_scatter_add_kernel<T, int64_t>(                          \
+        *x_grad, dim_val, shallow_copied_inds, values_grad, true, dev_ctx);  \
+  }                                                                          \
+  template <typename T, typename Context, EnableIfInteger<T> = 0>            \
+  void OpType##WithIndexGradKernel(const Context& dev_ctx,                   \
+                                   const DenseTensor& x,                     \
+                                   const DenseTensor& values,                \
+                                   const DenseTensor& indices,               \
+                                   const DenseTensor& values_grad,           \
+                                   const Scalar& dim,                        \
+                                   bool keepdim,                             \
+                                   DenseTensor* x_grad) {                    \
+    std::string dtype_name = phi::DataTypeToString(values.dtype());          \
+    PADDLE_ENFORCE_EQ(                                                       \
+        0,                                                                   \
+        1,                                                                   \
+        phi::errors::InvalidArgument(                                        \
+            "Integer type '%s' is not allowed to have stop_gradient=False.", \
+            dtype_name.c_str()));                                            \
+  }
+
+DEFINE_WITH_INDEX_GRAD_KERNEL(Max)
+DEFINE_WITH_INDEX_GRAD_KERNEL(Min)
+
+#undef DEFINE_WITH_INDEX_GRAD_KERNEL
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(max_with_index_grad,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::MaxWithIndexGradKernel,
+                   float,
+                   double,
+                   uint8_t,
+                   int,
+                   int16_t,
+                   int64_t,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}
+
+PD_REGISTER_KERNEL(min_with_index_grad,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::MinWithIndexGradKernel,
+                   float,
+                   double,
+                   uint8_t,
+                   int,
+                   int16_t,
+                   int64_t,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}