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

Save fused_attention op memory when dropout_rate = 0.0 #48902

Merged
merged 4 commits into from
Dec 13, 2022
Merged

Save fused_attention op memory when dropout_rate = 0.0 #48902

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
8de3f10
save fused_attention memory when dropout_rate = 0.0
sneaxiy Dec 8, 2022
cda1d66
add ut
sneaxiy Dec 12, 2022
9ca6460
fix ut bug
sneaxiy Dec 12, 2022
f802916
fix fused_layernorm_residual_dropout_bias_test.cu
sneaxiy Dec 13, 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
20 changes: 11 additions & 9 deletions paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_attention_fwd_func.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -572,15 +572,17 @@ fused_attention_dygraph_function(
egr::EagerUtils::CheckAndRetainGrad(SoftmaxOut);
grad_node->SetGradOutMeta(SoftmaxOut, 19);

auto AttnDropoutOut_accumulation_node =
std::make_shared<egr::GradNodeAccumulation>(
p_autograd_AttnDropoutOut);
egr::EagerUtils::SetOutRankWithSlot(p_autograd_AttnDropoutOut, 0);
egr::EagerUtils::SetHistory(p_autograd_AttnDropoutOut,
AttnDropoutOut_accumulation_node);
AttnDropoutOut_accumulation_node->SetGradInMeta(AttnDropoutOut, 0);
egr::EagerUtils::CheckAndRetainGrad(AttnDropoutOut);
grad_node->SetGradOutMeta(AttnDropoutOut, 20);
if (AttnDropoutOut.initialized()) {
auto AttnDropoutOut_accumulation_node =
std::make_shared<egr::GradNodeAccumulation>(
p_autograd_AttnDropoutOut);
egr::EagerUtils::SetOutRankWithSlot(p_autograd_AttnDropoutOut, 0);
egr::EagerUtils::SetHistory(p_autograd_AttnDropoutOut,
AttnDropoutOut_accumulation_node);
AttnDropoutOut_accumulation_node->SetGradInMeta(AttnDropoutOut, 0);
egr::EagerUtils::CheckAndRetainGrad(AttnDropoutOut);
grad_node->SetGradOutMeta(AttnDropoutOut, 20);
}

auto FMHAOut_accumulation_node =
std::make_shared<egr::GradNodeAccumulation>(p_autograd_FMHAOut);
Expand Down
2 changes: 1 addition & 1 deletion 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 @@ -476,7 +476,7 @@ class fused_attentionGradNodeCompat : public egr::GradNodeBase {
SoftmaxOut_ = egr::TensorWrapper(SoftmaxOut, false);
}
void SetTensorWrapperSrcMask(const paddle::experimental::Tensor& SrcMask) {
SrcMask_ = egr::TensorWrapper(SrcMask, false);
SrcMask_ = egr::TensorWrapper(SrcMask, true);
}
void SetTensorWrapperSrcMaskOut(
const paddle::experimental::Tensor& SrcMaskOut) {
Expand Down
6 changes: 3 additions & 3 deletions paddle/fluid/operators/fused/fmha_ref.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,7 +104,6 @@ class FMHARef {
T* qk_out_data = qk_out_tensor->data<T>();
T* qktv_out_data = qktv_out_tensor->data<T>();
T* softmax_out_data = softmax_out_tensor->data<T>();
T* dropout_out_data = dropout_out_tensor->data<T>();
T* fmha_out_data = fmha_out_tensor->data<T>();

auto out_seq_len = seq_len_;
Expand Down Expand Up @@ -221,6 +220,7 @@ class FMHARef {
dropout_mask_out_tensor,
dropout_out_tensor,
false);
T* dropout_out_data = dropout_out_tensor->data<T>();
blas.BatchedGEMM(transA,
transB,
gemm_m,
Expand Down Expand Up @@ -464,8 +464,6 @@ class FMHARef {

const T* softmax_out_data = softmax_out_tensor.data<T>();
T* softmax_out_grad_data = softmax_out_grad_tensor->data<T>();
const T* dropout_out_data = dropout_out_tensor.data<T>();
T* dropout_out_grad_data = dropout_out_grad_tensor->data<T>();
T* qktv_out_grad_data = qktv_out_grad_tensor->data<T>();

// transpose bw
Expand All @@ -487,6 +485,7 @@ class FMHARef {
int64_t stride_b = gemm_k * gemm_n;
// bw: dy = x^t * dout
if (dropout_param_.dropout_prob_) {
const T* dropout_out_data = dropout_out_tensor.data<T>();
blas.BatchedGEMM(transA,
transB,
gemm_m,
Expand Down Expand Up @@ -524,6 +523,7 @@ class FMHARef {
stride_a = gemm_m * gemm_k;
stride_b = gemm_k * gemm_n;
if (dropout_param_.dropout_prob_) {
T* dropout_out_grad_data = dropout_out_grad_tensor->data<T>();
blas.BatchedGEMM(transA,
transB,
gemm_m,
Expand Down
9 changes: 6 additions & 3 deletions 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 @@ -547,8 +547,10 @@ class FusedAttentionGradOp : public framework::OperatorWithKernel {
ctx->GetInputDim("QKOut"));
ctx->SetOutputDim(framework::GradVarName("SoftmaxOut"),
ctx->GetInputDim("SoftmaxOut"));
ctx->SetOutputDim(framework::GradVarName("AttnDropoutOut"),
ctx->GetInputDim("AttnDropoutOut"));
if (ctx->HasOutput(framework::GradVarName("AttnDropoutOut"))) {
ctx->SetOutputDim(framework::GradVarName("AttnDropoutOut"),
ctx->GetInputDim("AttnDropoutOut"));
}

if (ctx->HasOutput(framework::GradVarName("SrcMaskOut"))) {
ctx->SetOutputDim(framework::GradVarName("SrcMaskOut"),
Expand Down Expand Up @@ -709,7 +711,8 @@ DECLARE_NO_NEED_BUFFER_VARS_INFERER(FusedAttentionGradNoNeedBufferInferer,
"QKVOut",
"QKOut",
"QKTVOut",
"OutLinearOut");
"OutLinearOut",
"SrcMask");

} // namespace operators
} // namespace paddle
Expand Down
49 changes: 33 additions & 16 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 @@ -123,6 +123,10 @@ class FusedAttentionOpKernel : public framework::OpKernel<T> {
const float ln_epsilon = ctx.Attr<float>("ln_epsilon");

float attn_dropout_rate = ctx.Attr<float>("attn_dropout_rate");
const bool has_attn_dropout = (attn_dropout_rate != 0.0f);
DropoutParam dropout_param2(ctx, 0);
const bool has_dropout = (dropout_param2.dropout_prob != 0.0f);

bool is_test_1 = ctx.Attr<bool>("is_test");
auto &dropout_implementation_1 =
ctx.Attr<std::string>("attn_dropout_implementation");
Expand Down Expand Up @@ -171,11 +175,16 @@ class FusedAttentionOpKernel : public framework::OpKernel<T> {
src_mask_out->numel() * sizeof(T));
auto *softmax_out_data = dev_ctx.template Alloc<T>(
softmax_out, softmax_out->numel() * sizeof(T));
auto *attn_dropout_mask_out_data = dev_ctx.template Alloc<uint8_t>(
attn_dropout_mask_out,
attn_dropout_mask_out->numel() * sizeof(uint8_t));
auto *attn_dropout_out_data = dev_ctx.template Alloc<T>(
attn_dropout_out, attn_dropout_out->numel() * sizeof(T));
auto *attn_dropout_mask_out_data =
has_attn_dropout ? dev_ctx.template Alloc<uint8_t>(
attn_dropout_mask_out,
attn_dropout_mask_out->numel() * sizeof(uint8_t))
: nullptr;
auto *attn_dropout_out_data =
has_attn_dropout
? dev_ctx.template Alloc<T>(attn_dropout_out,
attn_dropout_out->numel() * sizeof(T))
: nullptr;
auto *fmha_out_data =
dev_ctx.template Alloc<T>(fmha_out, fmha_out->numel() * sizeof(T));

Expand All @@ -187,8 +196,11 @@ class FusedAttentionOpKernel : public framework::OpKernel<T> {
out_linear_out, out_linear_out->numel() * sizeof(T));

// get data ptr for bias+dropout+residual+layernorm
auto *dropout_mask_out_data = dev_ctx.template Alloc<uint8_t>(
dropout_mask_out, dropout_mask_out->numel() * sizeof(uint8_t));
auto *dropout_mask_out_data =
has_dropout
? dev_ctx.template Alloc<uint8_t>(
dropout_mask_out, dropout_mask_out->numel() * sizeof(uint8_t))
: nullptr;
auto *final_out_data =
dev_ctx.template Alloc<T>(out, out->numel() * sizeof(T));

Expand Down Expand Up @@ -248,7 +260,6 @@ class FusedAttentionOpKernel : public framework::OpKernel<T> {
input_size,
output_size,
false);
DropoutParam dropout_param2(ctx, 0);
FusedDropoutLayerNormHelper<T, uint8_t> fused_dropout_layernorm_helper(
ctx.cuda_device_context(),
bsz_seq,
Expand Down Expand Up @@ -369,7 +380,11 @@ class FusedAttentionGradKernel : public framework::OpKernel<T> {
const float epsilon = ctx.Attr<float>("epsilon");
const float ln2epsilon = ctx.Attr<float>("ln_epsilon");

float attn_dropout_prob = ctx.Attr<float>("attn_dropout_rate");
const float attn_dropout_prob = ctx.Attr<float>("attn_dropout_rate");
const bool has_attn_dropout = (attn_dropout_prob != 0.0f);
DropoutParam dropout_param2(ctx, 0);
const bool has_dropout = (dropout_param2.dropout_prob != 0.0f);

auto &dev_ctx = ctx.template device_context<phi::GPUContext>();
bool is_test_1 = ctx.Attr<bool>("is_test");
auto &dropout_implementation_1 =
Expand Down Expand Up @@ -400,7 +415,6 @@ class FusedAttentionGradKernel : public framework::OpKernel<T> {
auto *qkv_bias = ctx.Input<phi::DenseTensor>("QKVBias");
auto *out_linear_weight = ctx.Input<phi::DenseTensor>("OutLinearW");
auto *out_linear_bias = ctx.Input<phi::DenseTensor>("OutLinearBias");
auto *src_mask_data = (src_mask == nullptr ? nullptr : src_mask->data<T>());
auto *qkv_weight_data = qkv_weight->data<T>();
auto *qkv_bias_data = (qkv_bias == nullptr) ? nullptr : qkv_bias->data<T>();
auto *out_linear_weight_data = out_linear_weight->data<T>();
Expand All @@ -426,7 +440,8 @@ class FusedAttentionGradKernel : public framework::OpKernel<T> {
auto *softmax_out_data = softmax_out->data<T>();
auto *src_mask_out_data =
(src_mask == nullptr) ? nullptr : src_mask_out->data<T>();
auto *dropout_mask_out_data = dropout_mask_out->data<uint8_t>();
auto *dropout_mask_out_data =
has_dropout ? dropout_mask_out->data<uint8_t>() : nullptr;

// output's grad
auto *d_x = ctx.Output<phi::DenseTensor>(framework::GradVarName("X"));
Expand Down Expand Up @@ -472,8 +487,11 @@ class FusedAttentionGradKernel : public framework::OpKernel<T> {
dev_ctx.template Alloc<T>(d_qk_out, d_qk_out->numel() * sizeof(T));
auto *d_softmax_out_data = dev_ctx.template Alloc<T>(
d_softmax_out, d_softmax_out->numel() * sizeof(T));
auto *d_attn_dropout_out_data = dev_ctx.template Alloc<T>(
d_attn_dropout_out, d_attn_dropout_out->numel() * sizeof(T));
auto *d_attn_dropout_out_data =
has_attn_dropout
? dev_ctx.template Alloc<T>(d_attn_dropout_out,
d_attn_dropout_out->numel() * sizeof(T))
: nullptr;
auto *d_src_mask_out_data =
(src_mask == nullptr)
? nullptr
Expand Down Expand Up @@ -573,7 +591,6 @@ class FusedAttentionGradKernel : public framework::OpKernel<T> {
input_size,
output_size,
compute_bias);
DropoutParam dropout_param2(ctx, 0);
FusedDropoutLayerNormHelper<T, uint8_t> fused_dropout_layernorm_helper(
ctx.cuda_device_context(),
bsz_seq,
Expand Down Expand Up @@ -633,7 +650,7 @@ class FusedAttentionGradKernel : public framework::OpKernel<T> {

if (qkv_bias != nullptr) {
fmha_ref_compute.ComputeBackward(*transpose_out_2,
src_mask,
has_attn_dropout ? src_mask : nullptr,
*softmax_out,
*attn_dropout_mask_out,
*attn_dropout_out,
Expand All @@ -650,7 +667,7 @@ class FusedAttentionGradKernel : public framework::OpKernel<T> {
d_qkv_bias_out);
} else {
fmha_ref_compute.ComputeBackward(*transpose_out_2,
src_mask,
has_attn_dropout ? src_mask : nullptr,
*softmax_out,
*attn_dropout_mask_out,
*attn_dropout_out,
Expand Down
Loading