[cherry pick] add op: fused_feedforward(backward) (#36730)

* add op: fused_feedforward(backward) (#35611)

这个PR是fused_feedforward反向的代码

相关kernel实现：fused_dropout_act_bias, fused_residual_dropout_bias, fused_layernorm_residual_dropout_bias

fused_feedforward是一个融合算子，该算子对transformer模型的feed forward层的算子进行融合和封装，使得前端只呈现一个接口，通过融合减少部分访存和kernel launch的时间，以此提升性能。

* Move fused_attention and fused_feedforward functional api path to incubate (#36704)

将 #35905 和 #35843 PR中新增的的python api接口移到incubate目录下。

Author: zhangkaihuo

paddle/fluid/operators/fused/CMakeLists.txt

@@ -80,10 +80,8 @@ if (WITH_GPU OR WITH_ROCM)
         nv_test(test_fused_dropout_act_bias SRCS fused_dropout_act_bias_test.cu DEPS tensor op_registry dropout_op layer_norm_op device_context generator memory)
         nv_test(test_fused_layernorm_residual_dropout_bias SRCS fused_layernorm_residual_dropout_bias_test.cu DEPS tensor op_registry dropout_op layer_norm_op device_context generator memory)
 
-
         op_library(fused_feedforward_op)
         file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(fused_feedforward);\n")
-
         # fused_attention_op
         op_library(fused_attention_op)
         file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(fused_attention);\n")

paddle/fluid/operators/fused/fused_feedforward_op.cc

@@ -206,9 +206,154 @@ class FusedFeedForwardOpMaker : public framework::OpProtoAndCheckerMaker {
   }
 };
 
+class FusedFeedForwardOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {
+    PADDLE_ENFORCE_EQ(ctx->Attrs().Get<bool>("dropout1_is_test"), false,
+                      platform::errors::InvalidArgument(
+                          "GradOp is only callable when is_test is false"));
+    PADDLE_ENFORCE_EQ(ctx->Attrs().Get<bool>("dropout2_is_test"), false,
+                      platform::errors::InvalidArgument(
+                          "GradOp is only callable when is_test is false"));
+    OP_INOUT_CHECK(ctx->HasInput("Dropout1Mask"), "Input", "Dropout1Mask",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Dropout2Mask"), "Input", "Dropout1Mask",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Linear1Out"), "Input", "Linear1Out",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Ln1Out"), "Input", "Ln1Out",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Dropout1Out"), "Input", "Dropout1Out",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Dropout2Out"), "Input", "Dropout2Out",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Linear1Weight"), "Input", "Linear1Weight",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Linear2Weight"), "Input", "Linear2Weight",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Ln1Mean"), "Input", "Ln1Mean",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Ln1Variance"), "Input", "Ln1Variance",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Ln2Mean"), "Input", "Ln2Mean",
+                   "FusedFeedForwardGrad");
+    OP_INOUT_CHECK(ctx->HasInput("Ln2Variance"), "Input", "Ln2Variance",
+                   "FusedFeedForwardGrad");
+
+    OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input",
+                   framework::GradVarName("Out"), "FusedFeedForwardGrad");
+
+    auto d_out_dim = ctx->GetInputDim(framework::GradVarName("Out"));
+    ctx->SetOutputDim(framework::GradVarName("X"), d_out_dim);
+    if (ctx->HasOutput(framework::GradVarName("Ln1Scale"))) {
+      ctx->SetOutputDim(framework::GradVarName("Ln1Scale"),
+                        ctx->GetInputDim("Ln1Scale"));
+    }
+    if (ctx->HasOutput(framework::GradVarName("Ln1Bias"))) {
+      ctx->SetOutputDim(framework::GradVarName("Ln1Bias"),
+                        ctx->GetInputDim("Ln1Bias"));
+    }
+    if (ctx->HasOutput(framework::GradVarName("Ln2Scale"))) {
+      ctx->SetOutputDim(framework::GradVarName("Ln2Scale"),
+                        ctx->GetInputDim("Ln2Scale"));
+    }
+    if (ctx->HasOutput(framework::GradVarName("Ln2Bias"))) {
+      ctx->SetOutputDim(framework::GradVarName("Ln2Bias"),
+                        ctx->GetInputDim("Ln2Bias"));
+    }
+    ctx->SetOutputDim(framework::GradVarName("Linear1Weight"),
+                      ctx->GetInputDim("Linear1Weight"));
+    if (ctx->HasOutput(framework::GradVarName("Linear1Bias"))) {
+      ctx->SetOutputDim(framework::GradVarName("Linear1Bias"),
+                        ctx->GetInputDim("Linear1Bias"));
+    }
+    ctx->SetOutputDim(framework::GradVarName("Linear2Weight"),
+                      ctx->GetInputDim("Linear2Weight"));
+    if (ctx->HasOutput(framework::GradVarName("Linear2Bias"))) {
+      ctx->SetOutputDim(framework::GradVarName("Linear2Bias"),
+                        ctx->GetInputDim("Linear2Bias"));
+    }
+  }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext &ctx) const override {
+    auto input = ctx.Input<Tensor>("X");
+    auto input_data_type = input->type();
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
+};
+
+template <typename T>
+class FusedFeedForwardOpGradMaker : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+ protected:
+  void Apply(GradOpPtr<T> op) const override {
+    op->SetType("fused_feedforward_grad");
+    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
+    op->SetInput("X", this->Input("X"));
+    op->SetInput("Linear1Weight", this->Input("Linear1Weight"));
+    op->SetInput("Linear1Bias", this->Input("Linear1Bias"));
+    op->SetInput("Linear2Weight", this->Input("Linear2Weight"));
+    op->SetInput("Ln1Scale", this->Input("Ln1Scale"));
+    op->SetInput("Ln1Bias", this->Input("Ln1Bias"));
+    op->SetInput("Ln2Scale", this->Input("Ln2Scale"));
+    op->SetInput("Ln2Bias", this->Input("Ln2Bias"));
+    op->SetInput("Dropout1Mask", this->Output("Dropout1Mask"));
+    op->SetInput("Dropout2Mask", this->Output("Dropout2Mask"));
+    op->SetInput("Linear1Out", this->Output("Linear1Out"));
+    op->SetInput("Ln1Out", this->Output("Ln1Out"));
+    op->SetInput("Ln1Mean", this->Output("Ln1Mean"));
+    op->SetInput("Ln1Variance", this->Output("Ln1Variance"));
+    op->SetInput("Ln2Mean", this->Output("Ln2Mean"));
+    op->SetInput("Ln2Variance", this->Output("Ln2Variance"));
+    op->SetInput("Dropout1Out", this->Output("Dropout1Out"));
+    op->SetInput("Dropout2Out", this->Output("Dropout2Out"));
+
+    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
+    op->SetOutput(framework::GradVarName("Ln1Scale"),
+                  this->InputGrad("Ln1Scale"));
+    op->SetOutput(framework::GradVarName("Ln1Bias"),
+                  this->InputGrad("Ln1Bias"));
+    op->SetOutput(framework::GradVarName("Ln2Scale"),
+                  this->InputGrad("Ln2Scale"));
+    op->SetOutput(framework::GradVarName("Ln2Bias"),
+                  this->InputGrad("Ln2Bias"));
+    op->SetOutput(framework::GradVarName("Linear1Weight"),
+                  this->InputGrad("Linear1Weight"));
+    op->SetOutput(framework::GradVarName("Linear1Bias"),
+                  this->InputGrad("Linear1Bias"));
+    op->SetOutput(framework::GradVarName("Linear2Weight"),
+                  this->InputGrad("Linear2Weight"));
+    if (this->HasInput("Linear2Bias")) {
+      op->SetInput("Linear2Bias", this->Input("Linear2Bias"));
+      op->SetOutput(framework::GradVarName("Linear2Bias"),
+                    this->InputGrad("Linear2Bias"));
+    }
+
+    op->SetAttrMap(this->Attrs());
+  }
+};
+
+template <typename T>
+class FusedFeedForwardOpDoubleGradMaker
+    : public framework::SingleGradOpMaker<T> {
+ public:
+  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
+
+ protected:
+  void Apply(GradOpPtr<T> grad_op) const override {}
+};
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(fused_feedforward, ops::FusedFeedForwardOp,
-                  ops::FusedFeedForwardOpMaker);
+                  ops::FusedFeedForwardOpMaker,
+                  ops::FusedFeedForwardOpGradMaker<paddle::framework::OpDesc>,
+                  ops::FusedFeedForwardOpGradMaker<paddle::imperative::OpBase>);
+REGISTER_OPERATOR(fused_feedforward_grad, ops::FusedFeedForwardOpGrad);

paddle/fluid/operators/fused/fused_feedforward_op.cu

@@ -171,6 +171,210 @@ class FusedFeedForwardKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename DeviceContext, typename T>
+class FusedFeedForwardGradKernel : public framework::OpKernel<T> {
+ public:
+  void MatMulGrad(const platform::CUDADeviceContext& ctx,
+                  const framework::Tensor& d_out, const framework::Tensor& a,
+                  const framework::Tensor& b, framework::Tensor* d_a,
+                  framework::Tensor* d_b) const {
+    auto blas = math::GetBlas<DeviceContext, T>(ctx);
+    auto a_2d = FoldInitDims(a);
+    auto b_2d = FoldInitDims(b);
+    auto mat_dim_a = math::CreateMatrixDescriptor(a_2d.dims(), 0, true);
+    auto mat_dim_b = math::CreateMatrixDescriptor(b_2d.dims(), 0, true);
+    auto mat_dim_dout = math::CreateMatrixDescriptor(d_out.dims(), 0, false);
+    T alpha = static_cast<T>(1.0);
+    blas.MatMul(d_out, mat_dim_dout, b, mat_dim_b, alpha, d_a, T(0));
+    blas.MatMul(a, mat_dim_a, d_out, mat_dim_dout, alpha, d_b, T(0));
+  }
+
+  void FFNGrad(
+      const framework::Tensor& d_out, const framework::Tensor& x,
+      const framework::Tensor& dropout1_mask,
+      const framework::Tensor& dropout2_mask,
+      const framework::Tensor& linear1_out, const framework::Tensor& ln1_out,
+      const framework::Tensor& dropout1_out,
+      const framework::Tensor& dropout2_out,
+      const framework::Tensor& linear1_weight,
+      const framework::Tensor* linear1_bias,
+      const framework::Tensor& linear2_weight,
+      const framework::Tensor* ln1_gamma, const framework::Tensor* ln1_beta,
+      const framework::Tensor& ln1_mean, const framework::Tensor& ln1_variance,
+      const framework::Tensor* ln2_gamma, const framework::Tensor* ln2_beta,
+      const framework::Tensor& ln2_mean, const framework::Tensor& ln2_variance,
+      framework::Tensor* d_x, framework::Tensor* d_linear1_weight,
+      framework::Tensor* d_linear1_bias, framework::Tensor* d_linear2_weight,
+      framework::Tensor* d_linear2_bias, framework::Tensor* d_ln1_gamma,
+      framework::Tensor* d_ln1_beta, framework::Tensor* d_ln2_gamma,
+      framework::Tensor* d_ln2_beta, const int bsz_seq, const int d_model,
+      const int dim_feedforward, const DropoutParam& dropout_param1,
+      const DropoutParam& dropout_param2, const std::string& act_method,
+      const bool pre_layer_norm, const float epsilon1, const float epsilon2,
+      const platform::CUDADeviceContext& ctx) const {
+    FusedDropoutLayerNormHelper<T, uint8_t> pre_layernorm_helper(
+        bsz_seq, d_model, epsilon1);
+    FusedDropoutHelper<T, uint8_t> fused_act_dropout_helper(
+        ctx, bsz_seq, dim_feedforward, dropout_param1);
+    FusedDropoutLayerNormHelper<T, uint8_t> fused_dropout_layernorm_helper(
+        ctx, bsz_seq, d_model, dropout_param2, epsilon2);
+
+    auto place = ctx.GetPlace();
+    using U = LayerNormParamType<T>;
+    const U* ln1_gamma_ptr =
+        ln1_gamma == nullptr ? nullptr : ln1_gamma->data<U>();
+    const U* ln1_beta_ptr = ln1_beta == nullptr ? nullptr : ln1_beta->data<U>();
+    const U* ln2_gamma_ptr =
+        ln2_gamma == nullptr ? nullptr : ln2_gamma->data<U>();
+    const U* ln2_beta_ptr = ln2_beta == nullptr ? nullptr : ln2_beta->data<U>();
+    const T* linear1_bias_ptr =
+        linear1_bias == nullptr ? nullptr : linear1_bias->data<T>();
+    T* d_linear1_bias_ptr =
+        d_linear1_bias == nullptr ? nullptr : d_linear1_bias->data<T>();
+    T* d_linear2_bias_ptr =
+        d_linear2_bias == nullptr ? nullptr : d_linear2_bias->data<T>();
+    U* d_ln1_gamma_ptr =
+        d_ln1_gamma == nullptr ? nullptr : d_ln1_gamma->data<U>();
+    U* d_ln1_beta_ptr = d_ln1_beta == nullptr ? nullptr : d_ln1_beta->data<U>();
+    U* d_ln2_gamma_ptr =
+        d_ln2_gamma == nullptr ? nullptr : d_ln2_gamma->data<U>();
+    U* d_ln2_beta_ptr = d_ln2_beta == nullptr ? nullptr : d_ln2_beta->data<U>();
+
+    framework::Tensor d_linear2_out, d_dropout2_out, d_residual;
+    d_linear2_out.mutable_data<T>({bsz_seq, d_model}, place);
+    d_dropout2_out.mutable_data<T>({bsz_seq, d_model}, place);
+    d_residual.mutable_data<T>({bsz_seq, d_model}, place);
+
+    if (pre_layer_norm) {
+      fused_dropout_layernorm_helper.ResidualDropoutBiasGrad(
+          ctx, d_out.data<T>(), dropout2_mask.data<uint8_t>(),
+          d_linear2_out.data<T>(), d_residual.data<T>(), d_linear2_bias_ptr);
+    } else {
+      fused_dropout_layernorm_helper.LayernormResidualDropoutBiasGrad(
+          ctx, d_out.data<T>(), dropout2_out.data<T>(),
+          dropout2_mask.data<uint8_t>(), ln2_gamma_ptr, ln2_mean.data<U>(),
+          ln2_variance.data<U>(), d_dropout2_out.data<T>(), d_ln2_gamma_ptr,
+          d_ln2_beta_ptr, d_linear2_out.data<T>(), d_linear2_bias_ptr,
+          d_residual.data<T>());
+    }
+
+    framework::Tensor d_dropout1_out;
+    d_dropout1_out.mutable_data<T>({bsz_seq, dim_feedforward}, place);
+    MatMulGrad(ctx, d_linear2_out, dropout1_out, linear2_weight,
+               &d_dropout1_out, d_linear2_weight);
+
+    framework::Tensor d_linear1_out;
+    d_linear1_out.mutable_data<T>({bsz_seq, dim_feedforward}, place);
+    fused_act_dropout_helper.DropoutActBiasGrad(
+        ctx, d_dropout1_out.data<T>(), linear1_out.data<T>(), linear1_bias_ptr,
+        dropout1_mask.data<uint8_t>(), d_linear1_out.data<T>(),
+        d_linear1_bias_ptr, act_method);
+
+    if (pre_layer_norm) {
+      framework::Tensor d_ln1_out;
+      d_ln1_out.mutable_data<T>({bsz_seq, d_model}, place);
+      MatMulGrad(ctx, d_linear1_out, ln1_out, linear1_weight, &d_ln1_out,
+                 d_linear1_weight);
+
+      pre_layernorm_helper.LayerNormGrad(ctx, d_ln1_out.data<T>(), x.data<T>(),
+                                         ln1_gamma_ptr, ln1_mean.data<U>(),
+                                         ln1_variance.data<U>(), d_x->data<T>(),
+                                         d_ln1_gamma_ptr, d_ln1_beta_ptr);
+    } else {
+      MatMulGrad(ctx, d_linear1_out, x, linear1_weight, d_x, d_linear1_weight);
+    }
+  }
+
+  void Compute(const framework::ExecutionContext& context) const override {
+    using U = LayerNormParamType<T>;
+    auto d_out =
+        *context.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto x = *context.Input<framework::Tensor>("X");
+    auto dropout1_mask = *context.Input<framework::Tensor>("Dropout1Mask");
+    auto dropout2_mask = *context.Input<framework::Tensor>("Dropout2Mask");
+    auto linear1_out = *context.Input<framework::Tensor>("Linear1Out");
+    auto ln1_out = *context.Input<framework::Tensor>("Ln1Out");
+    auto dropout1_out = *context.Input<framework::Tensor>("Dropout1Out");
+    auto dropout2_out = *context.Input<framework::Tensor>("Dropout2Out");
+    auto linear1_weight = *context.Input<framework::Tensor>("Linear1Weight");
+    auto* linear1_bias = context.Input<framework::Tensor>("Linear1Bias");
+    auto linear2_weight = *context.Input<framework::Tensor>("Linear2Weight");
+    auto ln1_mean = *context.Input<framework::Tensor>("Ln1Mean");
+    auto ln1_variance = *context.Input<framework::Tensor>("Ln1Variance");
+    auto* ln1_scale = context.Input<framework::Tensor>("Ln1Scale");
+    auto* ln1_bias = context.Input<framework::Tensor>("Ln1Bias");
+    auto ln2_mean = *context.Input<framework::Tensor>("Ln2Mean");
+    auto ln2_variance = *context.Input<framework::Tensor>("Ln2Variance");
+    auto* ln2_scale = context.Input<framework::Tensor>("Ln2Scale");
+    auto* ln2_bias = context.Input<framework::Tensor>("Ln2Bias");
+
+    auto* d_x = context.Output<framework::Tensor>(framework::GradVarName("X"));
+    auto* d_ln1_scale =
+        context.Output<framework::Tensor>(framework::GradVarName("Ln1Scale"));
+    auto* d_ln1_bias =
+        context.Output<framework::Tensor>(framework::GradVarName("Ln1Bias"));
+    auto* d_ln2_scale =
+        context.Output<framework::Tensor>(framework::GradVarName("Ln2Scale"));
+    auto* d_ln2_bias =
+        context.Output<framework::Tensor>(framework::GradVarName("Ln2Bias"));
+    auto* d_linear1_weight = context.Output<framework::Tensor>(
+        framework::GradVarName("Linear1Weight"));
+    auto* d_linear1_bias = context.Output<framework::Tensor>(
+        framework::GradVarName("Linear1Bias"));
+    auto* d_linear2_weight = context.Output<framework::Tensor>(
+        framework::GradVarName("Linear2Weight"));
+    auto* d_linear2_bias = context.Output<framework::Tensor>(
+        framework::GradVarName("Linear2Bias"));
+
+    const float epsilon1 = context.Attr<float>("ln1_epsilon");
+    const float epsilon2 = context.Attr<float>("ln2_epsilon");
+    const bool pre_layer_norm = context.Attr<bool>("pre_layer_norm");
+    const std::string act_method = context.Attr<std::string>("act_method");
+    DropoutParam dropout_param1(context, 1);
+    DropoutParam dropout_param2(context, 2);
+
+    auto place = context.GetPlace();
+    d_x->mutable_data<T>(place);
+    if (d_ln1_scale) {
+      d_ln1_scale->mutable_data<U>(place);
+    }
+    if (d_ln1_bias) {
+      d_ln1_bias->mutable_data<U>(place);
+    }
+    if (d_ln2_scale) {
+      d_ln2_scale->mutable_data<U>(place);
+    }
+    if (d_ln2_bias) {
+      d_ln2_bias->mutable_data<U>(place);
+    }
+    if (d_linear1_bias) {
+      d_linear1_bias->mutable_data<T>(place);
+    }
+    if (d_linear2_bias) {
+      d_linear2_bias->mutable_data<T>(place);
+    }
+    d_linear1_weight->mutable_data<T>(place);
+    d_linear2_weight->mutable_data<T>(place);
+
+    auto x_dim = x.dims();
+    auto mat_dim_x =
+        math::CreateMatrixDescriptor(RowMatrixFromVector(x_dim), 0, false);
+
+    auto linear1_weight_dim = linear1_weight.dims();
+    int d_model = linear1_weight_dim[0];
+    int dim_feedforward = linear1_weight_dim[linear1_weight_dim.size() - 1];
+    int bsz_seq = mat_dim_x.batch_size_ * mat_dim_x.height_;
+
+    FFNGrad(d_out, x, dropout1_mask, dropout2_mask, linear1_out, ln1_out,
+            dropout1_out, dropout2_out, linear1_weight, linear1_bias,
+            linear2_weight, ln1_scale, ln1_bias, ln1_mean, ln1_variance,
+            ln2_scale, ln2_bias, ln2_mean, ln2_variance, d_x, d_linear1_weight,
+            d_linear1_bias, d_linear2_weight, d_linear2_bias, d_ln1_scale,
+            d_ln1_bias, d_ln2_scale, d_ln2_bias, bsz_seq, d_model,
+            dim_feedforward, dropout_param1, dropout_param2, act_method,
+            pre_layer_norm, epsilon1, epsilon2, context.cuda_device_context());
+  }
+};
 }  // namespace operators
 }  // namespace paddle
 
@@ -181,3 +385,10 @@ REGISTER_OP_CUDA_KERNEL(
     ops::FusedFeedForwardKernel<paddle::platform::CUDADeviceContext, double>,
     ops::FusedFeedForwardKernel<paddle::platform::CUDADeviceContext,
                                 paddle::platform::float16>);
+REGISTER_OP_CUDA_KERNEL(
+    fused_feedforward_grad,
+    ops::FusedFeedForwardGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::FusedFeedForwardGradKernel<paddle::platform::CUDADeviceContext,
+                                    double>,
+    ops::FusedFeedForwardGradKernel<paddle::platform::CUDADeviceContext,
+                                    paddle::platform::float16>);

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -91,7 +91,6 @@ foreach(TEST_OP ${MIXED_DIST_TEST_OPS})
 endforeach()
 
 if(NOT WITH_GPU)
-
     LIST(REMOVE_ITEM TEST_OPS test_fused_feedforward_op)
     LIST(REMOVE_ITEM TEST_OPS test_fused_attention_op)
 endif()