Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Refine the memory usage of fused_attention and fused_feedforward ops #47236

Merged
merged 5 commits into from
Oct 26, 2022
Merged

Refine the memory usage of fused_attention and fused_feedforward ops #47236

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
a346625
fix fused_attention fused_feedforward
sneaxiy Oct 20, 2022
db2ffed
fix ci
sneaxiy Oct 21, 2022
ba54aa5
fix ci
sneaxiy Oct 21, 2022
f3ef3e0
fix ci PADDLE_GET_CONST
sneaxiy Oct 21, 2022
98bdad1
fix ci ut
sneaxiy Oct 24, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 25 additions & 6 deletions paddle/fluid/eager/api/manual/fluid_manual/nodes/nodes.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,23 @@
#include "paddle/fluid/eager/tensor_wrapper.h"
#include "paddle/fluid/imperative/tracer.h"

template <typename T>
const T& GetAttrWithDefault(
const paddle::framework::AttributeMap& attrs,
const paddle::framework::AttributeMap& default_attrs,
const std::string& name) {
auto iter1 = attrs.find(name);
if (iter1 != attrs.end()) {
return PADDLE_GET_CONST(T, iter1->second);
}
auto iter2 = default_attrs.find(name);
if (iter2 != default_attrs.end()) {
return PADDLE_GET_CONST(T, iter2->second);
}
PADDLE_THROW(
phi::errors::InvalidArgument("Attribute(%s) cannot be found.", name));
}

class fused_gate_attentionGradNodeCompat : public egr::GradNodeBase {
public:
fused_gate_attentionGradNodeCompat() : egr::GradNodeBase() {
Expand Down Expand Up @@ -240,7 +257,9 @@ class fused_feedforwardGradNodeCompat : public egr::GradNodeBase {
}
void SetTensorWrapperDropout2Out(
const paddle::experimental::Tensor& Dropout2Out) {
Dropout2Out_ = egr::TensorWrapper(Dropout2Out, false);
auto pre_layer_norm = GetAttrWithDefault<bool>(
attr_map_, default_attr_map_, "pre_layer_norm");
Dropout2Out_ = egr::TensorWrapper(Dropout2Out, pre_layer_norm);
}
void SetTensorWrapperLinear1Bias(
const paddle::experimental::Tensor& Linear1Bias) {
Expand Down Expand Up @@ -427,27 +446,27 @@ class fused_attentionGradNodeCompat : public egr::GradNodeBase {
}
void SetTensorWrapperOutLinearOut(
const paddle::experimental::Tensor& OutLinearOut) {
OutLinearOut_ = egr::TensorWrapper(OutLinearOut, false);
OutLinearOut_ = egr::TensorWrapper(OutLinearOut, true);
}
void SetTensorWrapperOutLinearW(
const paddle::experimental::Tensor& OutLinearW) {
OutLinearW_ = egr::TensorWrapper(OutLinearW, false);
}
void SetTensorWrapperQKOut(const paddle::experimental::Tensor& QKOut) {
QKOut_ = egr::TensorWrapper(QKOut, false);
QKOut_ = egr::TensorWrapper(QKOut, true);
}
void SetTensorWrapperQKTVOut(const paddle::experimental::Tensor& QKTVOut) {
QKTVOut_ = egr::TensorWrapper(QKTVOut, false);
QKTVOut_ = egr::TensorWrapper(QKTVOut, true);
}
void SetTensorWrapperQKVBias(const paddle::experimental::Tensor& QKVBias) {
QKVBias_ = egr::TensorWrapper(QKVBias, false);
}
void SetTensorWrapperQKVBiasOut(
const paddle::experimental::Tensor& QKVBiasOut) {
QKVBiasOut_ = egr::TensorWrapper(QKVBiasOut, false);
QKVBiasOut_ = egr::TensorWrapper(QKVBiasOut, true);
}
void SetTensorWrapperQKVOut(const paddle::experimental::Tensor& QKVOut) {
QKVOut_ = egr::TensorWrapper(QKVOut, false);
QKVOut_ = egr::TensorWrapper(QKVOut, true);
}
void SetTensorWrapperQKVW(const paddle::experimental::Tensor& QKVW) {
QKVW_ = egr::TensorWrapper(QKVW, false);
Expand Down
11 changes: 10 additions & 1 deletion paddle/fluid/operators/fused/fused_attention_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -704,6 +704,13 @@ class FusedAttentionGradOpMaker : public framework::SingleGradOpMaker<T> {
}
};

DECLARE_NO_NEED_BUFFER_VARS_INFERER(FusedAttentionGradNoNeedBufferInferer,
"QKVBiasOut",
"QKVOut",
"QKOut",
"QKTVOut",
"OutLinearOut");

} // namespace operators
} // namespace paddle

Expand All @@ -713,7 +720,9 @@ REGISTER_OPERATOR(fused_attention,
ops::FusedAttentionOpMaker,
ops::FusedAttentionGradOpMaker<paddle::framework::OpDesc>,
ops::FusedAttentionGradOpMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(fused_attention_grad, ops::FusedAttentionGradOp);
REGISTER_OPERATOR(fused_attention_grad,
ops::FusedAttentionGradOp,
ops::FusedAttentionGradNoNeedBufferInferer);

REGISTER_OP_VERSION(fused_attention)
.AddCheckpoint(
Expand Down
5 changes: 0 additions & 5 deletions paddle/fluid/operators/fused/fused_attention_op.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -416,26 +416,21 @@ class FusedAttentionGradKernel : public framework::OpKernel<T> {
auto *fmha_out = ctx.Input<phi::DenseTensor>("FMHAOut");
auto *transpose_out_2 = ctx.Input<phi::DenseTensor>("TransposeOut2");
auto *qk_out = ctx.Input<phi::DenseTensor>("QKOut");
auto *qktv_out = ctx.Input<phi::DenseTensor>("QKTVOut");
auto *softmax_out = ctx.Input<phi::DenseTensor>("SoftmaxOut");
auto *attn_dropout_mask_out =
ctx.Input<phi::DenseTensor>("AttnDropoutMaskOut");
auto *attn_dropout_out = ctx.Input<phi::DenseTensor>("AttnDropoutOut");
auto *src_mask_out = ctx.Input<phi::DenseTensor>("SrcMaskOut");
auto *out_linear_out = ctx.Input<phi::DenseTensor>("OutLinearOut");
auto *ln_2_mean = ctx.Input<phi::DenseTensor>("Ln2Mean");
auto *ln_2_var = ctx.Input<phi::DenseTensor>("Ln2Variance");
auto *dropout_mask_out = ctx.Input<phi::DenseTensor>("DropoutMaskOut");
auto *bias_dropout_residual_out =
ctx.Input<phi::DenseTensor>("BiasDropoutResidualOut");
auto *fmha_out_data = fmha_out->data<T>();
auto *transpose_out_2_data = transpose_out_2->data<T>();
auto *qk_out_data = qk_out->data<T>();
auto *qktv_out_data = qktv_out->data<T>();
auto *softmax_out_data = softmax_out->data<T>();
auto *src_mask_out_data =
(src_mask == nullptr) ? nullptr : src_mask_out->data<T>();
auto *out_linear_out_data = out_linear_out->data<T>();
auto *dropout_mask_out_data = dropout_mask_out->data<uint8_t>();

// output's grad
Expand Down
14 changes: 9 additions & 5 deletions paddle/fluid/operators/fused/fused_feedforward_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -276,10 +276,12 @@ class FusedFeedForwardOpGrad : public framework::OperatorWithKernel {
"Input",
"Dropout1Out",
"FusedFeedForwardGrad");
OP_INOUT_CHECK(ctx->HasInput("Dropout2Out"),
"Input",
"Dropout2Out",
"FusedFeedForwardGrad");
if (!pre_layer_norm) {
OP_INOUT_CHECK(ctx->HasInput("Dropout2Out"),
"Input",
"Dropout2Out",
"FusedFeedForwardGrad");
}
OP_INOUT_CHECK(ctx->HasInput("Linear1Weight"),
"Input",
"Linear1Weight",
Expand Down Expand Up @@ -368,10 +370,12 @@ class FusedFeedForwardOpGradMaker : public framework::SingleGradOpMaker<T> {
op->SetInput("Dropout2Mask", this->Output("Dropout2Mask"));
op->SetInput("Linear1Out", this->Output("Linear1Out"));
op->SetInput("Dropout1Out", this->Output("Dropout1Out"));
op->SetInput("Dropout2Out", this->Output("Dropout2Out"));

op->SetAttrMap(this->Attrs());
bool pre_layer_norm = PADDLE_GET_CONST(bool, op->GetAttr("pre_layer_norm"));
if (!pre_layer_norm) {
op->SetInput("Dropout2Out", this->Output("Dropout2Out"));
}

op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
if (pre_layer_norm) {
Expand Down
6 changes: 3 additions & 3 deletions paddle/fluid/operators/fused/fused_feedforward_op.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -337,7 +337,7 @@ class FusedFeedForwardGradKernel : public framework::OpKernel<T> {
const phi::DenseTensor& linear1_out,
const phi::DenseTensor* ln1_out,
const phi::DenseTensor& dropout1_out,
const phi::DenseTensor& dropout2_out,
const phi::DenseTensor* dropout2_out,
const phi::DenseTensor& linear1_weight,
const phi::DenseTensor* linear1_bias,
const phi::DenseTensor& linear2_weight,
Expand Down Expand Up @@ -420,7 +420,7 @@ class FusedFeedForwardGradKernel : public framework::OpKernel<T> {
fused_dropout_layernorm_helper.LayernormResidualDropoutBiasGrad(
ctx,
d_out.data<T>(),
dropout2_out.data<T>(),
dropout2_out->data<T>(),
dropout2_mask.data<uint8_t>(),
ln2_gamma_ptr,
ln2_mean->data<U>(),
Expand Down Expand Up @@ -504,7 +504,7 @@ class FusedFeedForwardGradKernel : public framework::OpKernel<T> {
auto* ln1_out =
pre_layer_norm ? context.Input<phi::DenseTensor>("Ln1Out") : nullptr;
auto dropout1_out = *context.Input<phi::DenseTensor>("Dropout1Out");
auto dropout2_out = *context.Input<phi::DenseTensor>("Dropout2Out");
auto* dropout2_out = context.Input<phi::DenseTensor>("Dropout2Out");
auto linear1_weight = *context.Input<phi::DenseTensor>("Linear1Weight");
auto* linear1_bias = context.Input<phi::DenseTensor>("Linear1Bias");
auto linear2_weight = *context.Input<phi::DenseTensor>("Linear2Weight");
Expand Down