support fuse layer norm grad for npu (#10614)

ShawnXuan · crazy-JiangDongHua · oneflow-ci-bot · web-flow · commit cb699cd17feb · 2025-01-23T15:39:04.000+08:00
Co-authored-by: JiangDongHua &lt;759421566@qq.com&gt;
Co-authored-by: oneflow-ci-bot &lt;ci-bot@oneflow.org&gt;
diff --git a/oneflow/core/autograd/gradient_funcs/layer_norm.cpp b/oneflow/core/autograd/gradient_funcs/layer_norm.cpp
@@ -18,6 +18,9 @@ limitations under the License.
 #include "oneflow/core/functional/functional.h"
 
 namespace oneflow {
+
+DEFINE_ENV_BOOL(ONEFLOW_USE_FUSE_LAYER_NORM_GRAD, false);
+
 namespace one {
 
 struct LayerNormCaptureState : public AutoGradCaptureState {
@@ -107,22 +110,36 @@ Maybe<void> LayerNorm::Apply(const LayerNormCaptureState* ctx, const TensorTuple
   std::shared_ptr<Tensor> mean = saved_tensors.at(ctx->mean_index);
   std::shared_ptr<Tensor> inv_variance = saved_tensors.at(ctx->inv_variance_index);
 
-  if (ctx->has_affine) {
-    // Use LayerNormParamGrad(Tensor dy, Tensor x, Tensor mean, Tensor inv_variance,
-    // Int64 begin_params_axis)
-    const auto& results =
-        JUST(functional::LayerNormParamGrad(dy, x, mean, inv_variance, begin_params_axis));
-    in_grads->at(1) = results->at(0);  // For gamma.
-    in_grads->at(2) = results->at(1);  // For beta.
-  }
-  if (ctx->x_requires_grad) {
-    if (ctx->scale) {
-      std::shared_ptr<Tensor> gamma = saved_tensors.at(ctx->gamma_index);
-      in_grads->at(0) = JUST(functional::LayerNormAffineGrad(dy, x, mean, inv_variance, gamma,
-                                                             begin_norm_axis, ctx->epsilon));
-    } else {
-      in_grads->at(0) =
-          JUST(functional::LayerNormGrad(dy, x, mean, inv_variance, begin_norm_axis, ctx->epsilon));
+  if (EnvBool<ONEFLOW_USE_FUSE_LAYER_NORM_GRAD>()) {
+    // just for npu
+    CHECK(ctx->has_affine) << "LayerNorm::Apply must has_affine for NPU GPT2 test";
+    if (ctx->x_requires_grad) {
+      if (ctx->scale) {
+        std::shared_ptr<Tensor> gamma = saved_tensors.at(ctx->gamma_index);
+        *in_grads = *JUST(functional::FuseLayerNormGrad(
+            dy, x, mean, inv_variance, gamma, begin_norm_axis, begin_params_axis, ctx->epsilon));
+      } else {
+        UNIMPLEMENTED();
+      }
+    }
+  } else {
+    if (ctx->has_affine) {
+      // Use LayerNormParamGrad(Tensor dy, Tensor x, Tensor mean, Tensor inv_variance,
+      // Int64 begin_params_axis)
+      const auto& results =
+          JUST(functional::LayerNormParamGrad(dy, x, mean, inv_variance, begin_params_axis));
+      in_grads->at(1) = results->at(0);  // For gamma.
+      in_grads->at(2) = results->at(1);  // For beta.
+    }
+    if (ctx->x_requires_grad) {
+      if (ctx->scale) {
+        std::shared_ptr<Tensor> gamma = saved_tensors.at(ctx->gamma_index);
+        in_grads->at(0) = JUST(functional::LayerNormAffineGrad(dy, x, mean, inv_variance, gamma,
+                                                               begin_norm_axis, ctx->epsilon));
+      } else {
+        in_grads->at(0) = JUST(
+            functional::LayerNormGrad(dy, x, mean, inv_variance, begin_norm_axis, ctx->epsilon));
+      }
     }
   }
   return Maybe<void>::Ok();
diff --git a/oneflow/core/functional/functional_api.yaml b/oneflow/core/functional/functional_api.yaml
@@ -1558,6 +1558,10 @@
   signature: "Tensor (Tensor dy, Tensor x, Tensor mean, Tensor inv_variance, Tensor gamma, Int64 begin_norm_axis, Double epsilon) => LayerNormAffineGrad"
   bind_python: False
 
+- name: "fuse_layer_norm_grad"
+  signature: "TensorTuple (Tensor dy, Tensor x, Tensor mean, Tensor inv_variance, Tensor gamma, Int64 begin_norm_axis, Int64 begin_params_axis, Double epsilon) => FuseLayerNormGrad"
+  bind_python: False
+
 - name: "layer_norm_param_grad"
   signature: "TensorTuple (Tensor dy, Tensor x, Tensor mean, Tensor inv_variance, Int64 begin_params_axis) => LayerNormParamGrad"
   bind_python: False
diff --git a/oneflow/core/functional/impl/nn_grad_functor.cpp b/oneflow/core/functional/impl/nn_grad_functor.cpp
@@ -983,6 +983,36 @@ class LayerNormAffineGradFunctor {
   std::shared_ptr<OpExpr> op_;
 };
 
+class FuseLayerNormGradFunctor {
+ public:
+  FuseLayerNormGradFunctor() {
+    op_ = CHECK_JUST(one::OpBuilder("fuse_layer_norm_grad")
+                         .Input("dy")
+                         .Input("x")
+                         .Input("mean")
+                         .Input("inv_variance")
+                         .Input("gamma")
+                         .Output("dx")
+                         .Output("gamma_diff")
+                         .Output("beta_diff")
+                         .Build());
+  }
+  Maybe<TensorTuple> operator()(const std::shared_ptr<one::Tensor>& dy,
+                                const std::shared_ptr<one::Tensor>& x,
+                                const std::shared_ptr<one::Tensor>& mean,
+                                const std::shared_ptr<one::Tensor>& inv_variance,
+                                const std::shared_ptr<one::Tensor>& gamma,
+                                const int64_t& begin_norm_axis, const int64_t& begin_params_axis,
+                                const double& epsilon) const {
+    auto& attrs = THREAD_CACHED_MUTABLE_ATTR_MAP("begin_norm_axis", "begin_params_axis", "epsilon");
+    attrs.SetAllAttrs(begin_norm_axis, begin_params_axis, epsilon);
+    return OpInterpUtil::Dispatch<TensorTuple>(*op_, {dy, x, mean, inv_variance, gamma}, attrs);
+  }
+
+ private:
+  std::shared_ptr<OpExpr> op_;
+};
+
 class LayerNormParamGradFunctor {
  public:
   LayerNormParamGradFunctor() {
@@ -1707,6 +1737,7 @@ ONEFLOW_FUNCTION_LIBRARY(m) {
   m.add_functor<impl::LayerNormGradFunctor>("LayerNormGrad");
   m.add_functor<impl::LayerNormAffineGradFunctor>("LayerNormAffineGrad");
   m.add_functor<impl::LayerNormParamGradFunctor>("LayerNormParamGrad");
+  m.add_functor<impl::FuseLayerNormGradFunctor>("FuseLayerNormGrad");
   m.add_functor<impl::GroupNormGradFunctor>("GroupNormGrad");
   m.add_functor<impl::GroupNormParamGradFunctor>("GroupNormParamGrad");
   m.add_functor<impl::BroadcastMatmulGradBFunctor>("BroadcastMatmulGradB");
diff --git a/oneflow/core/job_rewriter/auto_mixed_precision.cpp b/oneflow/core/job_rewriter/auto_mixed_precision.cpp
@@ -359,6 +359,8 @@ REGISTER_NO_CAST_REGISTRY("layer_norm_grad", "mean", 0)
 REGISTER_NO_CAST_REGISTRY("layer_norm_grad", "inv_variance", 0)
 REGISTER_NO_CAST_REGISTRY("layer_norm_param_grad", "mean", 0)
 REGISTER_NO_CAST_REGISTRY("layer_norm_param_grad", "inv_variance", 0)
+REGISTER_NO_CAST_REGISTRY("fuse_layer_norm_grad", "mean", 0)
+REGISTER_NO_CAST_REGISTRY("fuse_layer_norm_grad", "inv_variance", 0)
 
 }  // namespace
 
diff --git a/oneflow/core/job_rewriter/auto_mixed_precision_lists.cpp b/oneflow/core/job_rewriter/auto_mixed_precision_lists.cpp
@@ -95,6 +95,7 @@ const AMPList& AutoMixedPrecisionLists::GrayList() {
                               "layer_norm",
                               "layer_norm_param_grad",
                               "layer_norm_grad",
+                              "fuse_layer_norm_grad",
                               "skip_layer_norm",
                               "rms_norm",
                               "rms_norm_grad",
diff --git a/oneflow/ir/include/OneFlow/OneFlowUserOps.td b/oneflow/ir/include/OneFlow/OneFlowUserOps.td
@@ -7071,6 +7071,35 @@ def OneFlow_LayerNormGradOp : OneFlow_BaseOp<"layer_norm_grad", [NoMemoryEffect,
   let has_data_type_infer_fn = 1;
 }
 
+def OneFlow_FuseLayerNormGradOp : OneFlow_BaseOp<"fuse_layer_norm_grad", [NoMemoryEffect, AttrSizedOperandSegments, DeclareOpInterfaceMethods<UserOpCompatibleInterface>]> {
+  let input = (ins
+    OneFlow_Tensor:$dy,
+    OneFlow_Tensor:$x,
+    OneFlow_Tensor:$mean,
+    OneFlow_Tensor:$inv_variance,
+    Optional<OneFlow_Tensor>:$gamma,
+    Optional<OneFlow_Tensor>:$_add_to_output
+  );
+  let output = (outs
+    OneFlow_Tensor:$dx,
+    OneFlow_Tensor:$gamma_diff,
+    OneFlow_Tensor:$beta_diff
+  );
+  let attrs = (ins
+    DefaultValuedAttr<SI64Attr, "0">:$begin_norm_axis,
+    DefaultValuedAttr<SI64Attr, "0">:$begin_params_axis,
+    DefaultValuedAttr<F64Attr, "0.">:$epsilon
+  );
+  let trait_attrs = (ins
+    DenseI32ArrayAttr:$operand_segment_sizes,
+    DenseI32ArrayAttr:$result_segment_sizes
+  );
+  let has_logical_tensor_desc_infer_fn = 1;
+  let has_physical_tensor_desc_infer_fn = 1;
+  let has_get_sbp_fn = 1;
+  let has_data_type_infer_fn = 1;
+}
+
 def OneFlow_LayerNormParamGradOp : OneFlow_BaseOp<"layer_norm_param_grad", [NoMemoryEffect, AttrSizedResultSegments, DeclareOpInterfaceMethods<UserOpCompatibleInterface>]> {
   let input = (ins
     OneFlow_Tensor:$dy,
diff --git a/oneflow/user/kernels/layer_norm_cpu_kernel.cpp b/oneflow/user/kernels/layer_norm_cpu_kernel.cpp
@@ -57,6 +57,26 @@ class LayerNormGradCpuKernel final : public user_op::OpKernel {
 REGISTER_LAYER_NORM_GRAD_CPU_KERNEL(float)
 REGISTER_LAYER_NORM_GRAD_CPU_KERNEL(double)
 
+template<typename T>
+class FuseLayerNormGradCpuKernel final : public user_op::OpKernel {
+ public:
+  FuseLayerNormGradCpuKernel() = default;
+  ~FuseLayerNormGradCpuKernel() = default;
+
+ private:
+  bool AlwaysComputeWhenAllOutputsEmpty() const override { return false; }
+  void Compute(user_op::KernelComputeContext* ctx) const override { TODO(); };
+};
+
+#define REGISTER_FUSE_LAYER_NORM_GRAD_CPU_KERNEL(dtype)               \
+  REGISTER_USER_KERNEL("fuse_layer_norm_grad")                        \
+      .SetCreateFn<LayerNormGradCpuKernel<dtype>>()                   \
+      .SetIsMatchedHob((user_op::HobDeviceType() == DeviceType::kCPU) \
+                       && (user_op::HobDataType("dy", 0) == GetDataType<dtype>::value));
+
+REGISTER_FUSE_LAYER_NORM_GRAD_CPU_KERNEL(float)
+REGISTER_FUSE_LAYER_NORM_GRAD_CPU_KERNEL(double)
+
 template<typename T>
 class LayerNormParamGradCpuKernel final : public user_op::OpKernel {
  public:
diff --git a/oneflow/user/ops/layer_norm_op.cpp b/oneflow/user/ops/layer_norm_op.cpp
@@ -268,4 +268,129 @@ oneflow::DataType InferBnParamDataType(const DataType x_data_type) {
   return Maybe<void>::Ok();
 }
 
+/* static */ Maybe<void> FuseLayerNormGradOp::InferLogicalTensorDesc(user_op::InferContext* ctx) {
+  const user_op::TensorDesc& dy = ctx->InputTensorDesc("dy", 0);
+  const user_op::TensorDesc& x = ctx->InputTensorDesc("x", 0);
+  const user_op::TensorDesc& mean = ctx->InputTensorDesc("mean", 0);
+  const user_op::TensorDesc& inv_variance = ctx->InputTensorDesc("inv_variance", 0);
+  user_op::TensorDesc* dx = ctx->MutOutputTensorDesc("dx", 0);
+  CHECK_EQ_OR_RETURN(dy.shape(), x.shape()) << "dy and x shapes should be equal.";
+  const int64_t begin_norm_axis = ctx->Attr<int64_t>("begin_norm_axis");
+  CHECK_GT_OR_RETURN(begin_norm_axis, 0) << "begin_norm_axis must be greater than 0.";
+  const Shape& bn_param_shape = InferBnParamShape(x.shape(), begin_norm_axis);
+  CHECK_EQ_OR_RETURN(mean.shape(), bn_param_shape) << "mean shape must match bn_param_shape.";
+  CHECK_EQ_OR_RETURN(inv_variance.shape(), bn_param_shape)
+      << "inv_variance shape must match bn_param_shape.";
+  dx->set_shape(dy.shape());
+  dx->set_is_dynamic(dy.is_dynamic());
+  if (ctx->has_input("_add_to_output", 0)) {
+    const auto& add_to_output = ctx->InputTensorDesc("_add_to_output", 0);
+    CHECK_EQ_OR_RETURN(add_to_output.shape(), dx->shape())
+        << "add_to_output shape must match dx shape.";
+  }
+
+  auto has_tensor = [ctx](const std::string& bn) -> bool {
+    bool ret = false;
+    for (const auto& t : ctx->inputs()) {
+      if (bn == t.first) { return true; }
+    }
+    for (const auto& t : ctx->outputs()) {
+      if (bn == t.first) { return true; }
+    }
+    return ret;
+  };
+  const int64_t begin_params_axis = ctx->Attr<int64_t>("begin_params_axis");
+  const bool has_beta_diff = has_tensor("beta_diff");
+  const bool has_gamma_diff = has_tensor("gamma_diff");
+  CHECK_GE_OR_RETURN(begin_params_axis, 1)
+      << "begin_params_axis must be greater than or equal to 1.";
+  CHECK_LT_OR_RETURN(begin_params_axis, dy.shape().NumAxes())
+      << "begin_params_axis must be less than the number of axes in dy shape.";
+  DimVector param_shape_dim_vec;
+  param_shape_dim_vec.insert(param_shape_dim_vec.end(),
+                             dy.shape().dim_vec().cbegin() + begin_params_axis,
+                             dy.shape().dim_vec().cend());
+  const Shape param_shape(param_shape_dim_vec);
+  if (has_beta_diff) {
+    user_op::TensorDesc* beta_diff = ctx->MutOutputTensorDesc("beta_diff", 0);
+    beta_diff->set_shape(param_shape);
+  }
+  if (has_gamma_diff) {
+    user_op::TensorDesc* gamma_diff = ctx->MutOutputTensorDesc("gamma_diff", 0);
+    gamma_diff->set_shape(param_shape);
+  }
+  return Maybe<void>::Ok();
+}
+
+/*static*/ Maybe<void> FuseLayerNormGradOp::InferPhysicalTensorDesc(user_op::InferContext* ctx) {
+  return InferLogicalTensorDesc(ctx);
+}
+
+/* static */ Maybe<void> FuseLayerNormGradOp::GetSbp(user_op::SbpContext* ctx) {
+  std::vector<user_op::OpArg> broadcast_args;
+  if (ctx->user_op_conf().has_input("gamma", 0)) { broadcast_args.emplace_back("gamma", 0); }
+  int64_t begin_norm_axis = ctx->Attr<int64_t>("begin_norm_axis");
+  int64_t begin_params_axis = ctx->Attr<int64_t>("begin_params_axis");
+  CHECK_EQ(begin_norm_axis, begin_params_axis)
+      << "begin_norm_axis and begin_params_axis must be equal, but got " << begin_norm_axis
+      << " and " << begin_params_axis;
+  for (int i = 0; i < begin_norm_axis; ++i) {
+    ctx->NewBuilder()
+        .Split(ctx->inputs(), i)
+        .Split(user_op::OpArg("dx", 0), i)
+        .PartialSum(user_op::OpArg("gamma_diff", 0))
+        .PartialSum(user_op::OpArg("beta_diff", 0))
+        .Broadcast(broadcast_args)
+        .Build();
+  }
+  return Maybe<void>::Ok();
+}
+
+/* static */ Maybe<void> FuseLayerNormGradOp::InferDataType(user_op::InferContext* ctx) {
+  const user_op::TensorDesc& dy = ctx->InputTensorDesc("dy", 0);
+  const user_op::TensorDesc& x = ctx->InputTensorDesc("x", 0);
+  CHECK_EQ_OR_RETURN(dy.data_type(), x.data_type())
+      << "InferDataType Failed. Expected " << DataType_Name(x.data_type()) << ", but got "
+      << DataType_Name(dy.data_type());
+  const user_op::TensorDesc& mean = ctx->InputTensorDesc("mean", 0);
+  const user_op::TensorDesc& inv_variance = ctx->InputTensorDesc("inv_variance", 0);
+  DataType bn_param_data_type = InferBnParamDataType(x.data_type());
+  CHECK_EQ_OR_RETURN(mean.data_type(), bn_param_data_type)
+      << "InferDataType Failed. Expected " << DataType_Name(bn_param_data_type) << ", but got "
+      << DataType_Name(mean.data_type());
+  CHECK_EQ_OR_RETURN(inv_variance.data_type(), bn_param_data_type)
+      << "InferDataType Failed. Expected " << DataType_Name(bn_param_data_type) << ", but got "
+      << DataType_Name(inv_variance.data_type());
+  user_op::TensorDesc* dx = ctx->MutOutputTensorDesc("dx", 0);
+  dx->set_data_type(dy.data_type());
+  if (ctx->has_input("_add_to_output", 0)) {
+    const auto& add_to_output = ctx->InputTensorDesc("_add_to_output", 0);
+    CHECK_EQ_OR_RETURN(add_to_output.data_type(), dx->data_type())
+        << "InferDataType Failed. Expected " << DataType_Name(dx->data_type()) << ", but got "
+        << DataType_Name(add_to_output.data_type());
+  }
+
+  auto has_tensor = [ctx](const std::string& bn) -> bool {
+    bool ret = false;
+    for (auto& t : ctx->inputs()) {
+      if (bn == t.first) { return true; }
+    }
+    for (auto& t : ctx->outputs()) {
+      if (bn == t.first) { return true; }
+    }
+    return ret;
+  };
+  const bool has_beta_diff = has_tensor("beta_diff");
+  const bool has_gamma_diff = has_tensor("gamma_diff");
+  if (has_beta_diff) {
+    user_op::TensorDesc* beta_diff = ctx->MutOutputTensorDesc("beta_diff", 0);
+    beta_diff->set_data_type(dy.data_type());
+  }
+  if (has_gamma_diff) {
+    user_op::TensorDesc* gamma_diff = ctx->MutOutputTensorDesc("gamma_diff", 0);
+    gamma_diff->set_data_type(dy.data_type());
+  }
+  return Maybe<void>::Ok();
+}
+
 }  // namespace oneflow
