test=document_fix (#73561)

Author: lshpku

paddle/phi/infermeta/ternary.cc

@@ -1630,43 +1630,6 @@ void MoeCombineInferMeta(const MetaTensor& x,
   y->set_dtype(x.dtype());
 }
 
-void MoeCombineNoWeightInferMeta(const MetaTensor& x,
-                                 const MetaTensor& scatter_index,
-                                 MetaTensor* y) {
-  auto x_dim = x.dims();
-  auto scatter_index_dim = scatter_index.dims();
-  PADDLE_ENFORCE_EQ(x_dim.size(),
-                    2,
-                    common::errors::InvalidArgument(
-                        "The dimensions of Input(x) must be 2, but "
-                        "received dimensions of Input(x) is [%d]",
-                        x_dim.size()));
-  PADDLE_ENFORCE_EQ(scatter_index_dim.size(),
-                    2,
-                    common::errors::InvalidArgument(
-                        "The dimensions of Input(scatter_index) must be 2, but "
-                        "received dimensions of Input(scatter_index) is [%d]",
-                        scatter_index_dim.size()));
-  PADDLE_ENFORCE_EQ(scatter_index.dtype(),
-                    phi::DataType::INT32,
-                    common::errors::InvalidArgument(
-                        "The input scatter_index type should be int32"
-                        "But received scatter_index type = %s",
-                        scatter_index.dtype()));
-  int64_t seqlen = scatter_index_dim[0];
-  int64_t k = scatter_index_dim[1];
-  int64_t hidden_size = x_dim[1];
-  PADDLE_ENFORCE_EQ(x_dim[0],
-                    seqlen * k,
-                    common::errors::InvalidArgument(
-                        "The upper dim of Input(x) [%d] must equal to "
-                        "the total size of Input(scatter_index) [%d].",
-                        x_dim[0],
-                        seqlen * k));
-  y->set_dims(phi::make_ddim({seqlen, hidden_size}));
-  y->set_dtype(x.dtype());
-}
-
 void MoeGateDispatchPartialNoSoftmaxTopKInferMeta(
     const MetaTensor& x,
     const MetaTensor& combine_weights,

paddle/phi/infermeta/ternary.h

@@ -274,10 +274,6 @@ void MoeCombineInferMeta(const MetaTensor& x,
                          const MetaTensor& scatter_index,
                          MetaTensor* y);
 
-void MoeCombineNoWeightInferMeta(const MetaTensor& x,
-                                 const MetaTensor& scatter_index,
-                                 MetaTensor* y);
-
 void MoeGateDispatchPartialNoSoftmaxTopKInferMeta(
     const MetaTensor& x,
     const MetaTensor& combine_weights,

paddle/phi/kernels/CMakeLists.txt

@@ -64,8 +64,6 @@ if(((WITH_GPU) AND (CUDA_VERSION VERSION_LESS 12.0))
     "legacy/gpu/expand_modality_expert_id_kernel.cu"
     "legacy/gpu/moe_combine_kernel.cu"
     "legacy/gpu/moe_combine_grad_kernel.cu"
-    "legacy/gpu/moe_combine_no_weight_kernel.cu"
-    "legacy/gpu/moe_combine_no_weight_grad_kernel.cu"
     "legacy/gpu/cal_aux_loss_kernel.cu"
     "legacy/gpu/cal_aux_loss_grad_kernel.cu"
     "legacy/gpu/ext_build_src_rank_and_local_expert_id_kernel.cu"

paddle/phi/kernels/legacy/gpu/moe_combine_no_weight_grad_kernel.cu

@@ -2320,17 +2320,6 @@
   kernel :
     func : moe_combine_grad
 
-- backward_op : moe_combine_no_weight_grad
-  forward : moe_combine_no_weight (Tensor x, Tensor scatter_index) -> Tensor(y)
-  args : (Tensor x, Tensor scatter_index, Tensor y_grad)
-  output : Tensor(x_grad)
-  infer_meta :
-    func : UnchangedInferMeta
-    param : [x]
-  kernel :
-    func : moe_combine_no_weight_grad
-  no_need_buffer : x
-
 - backward_op : moe_gate_dispatch_grad
   forward : moe_gate_dispatch (Tensor x, Tensor gate_logits, Tensor corr_bias, int64_t k, int64_t capacity, bool use_pad) -> Tensor(y), Tensor(combine_weights), Tensor(scatter_index), Tensor(expert_offset), Tensor(expert_id)
   args : (Tensor combine_weights, Tensor scatter_index, Tensor expert_id, Tensor y_grad, Tensor combine_weights_grad, int64_t k, int64_t capacity, bool use_pad)

paddle/phi/kernels/legacy/gpu/moe_combine_no_weight_kernel.cu

@@ -3678,16 +3678,6 @@
     data_type : x
   backward : moe_combine_grad
 
-- op : moe_combine_no_weight
-  args : (Tensor x, Tensor scatter_index)
-  output : Tensor(y)
-  infer_meta :
-    func : MoeCombineNoWeightInferMeta
-  kernel :
-    func : moe_combine_no_weight
-    data_type : x
-  backward : moe_combine_no_weight_grad
-
 - op : moe_gate_dispatch
   args : (Tensor x, Tensor gate_logits, Tensor corr_bias, int64_t k, int64_t capacity, bool use_pad)
   output : Tensor(y), Tensor(combine_weights), Tensor(scatter_index), Tensor(expert_offset), Tensor(expert_id)

paddle/phi/ops/yaml/backward.yaml

@@ -60,7 +60,6 @@
 from .int_bincount import int_bincount
 from .masked_multihead_attention import masked_multihead_attention
 from .moe_combine import moe_combine
-from .moe_combine_no_weight import moe_combine_no_weight
 from .moe_gate_dispatch import moe_gate_dispatch
 from .moe_gate_dispatch_partial_nosoftmaxtopk import (
     moe_gate_dispatch_partial_nosoftmaxtopk,

paddle/phi/ops/yaml/ops.yaml

@@ -513,7 +513,6 @@ if(NOT WITH_GPU
     test_incubate_fused_rmsnorm_ext
     test_incubate_int_bincount
     test_incubate_moe_combine
-    test_incubate_moe_combine_no_weight
     test_incubate_moe_gate_dispatch_partial_nosoftmaxtopk
     test_incubate_moe_gate_dispatch_w_permute_bwd
     test_incubate_moe_gate_dispatch_w_permute