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

[Dygraph] Fix Perf of FusedFeedForward and FusedAttention with AllReduce #46780

Merged
merged 2 commits into from
Oct 9, 2022
Merged

[Dygraph] Fix Perf of FusedFeedForward and FusedAttention with AllReduce #46780

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
102ab57
fix perf of fuse op with allreduce
haohongxiang Oct 8, 2022
b5e32b7
update
haohongxiang Oct 8, 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
6 changes: 4 additions & 2 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 @@ -30,7 +30,7 @@ limitations under the License. */
#include "paddle/phi/kernels/funcs/math_function.h"

#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
#include "paddle/fluid/distributed/collective/ProcessGroup.h"
#include "paddle/fluid/distributed/collective/ProcessGroupNCCL.h"
#include "paddle/fluid/platform/collective_helper.h"
#include "paddle/fluid/platform/device/gpu/nccl_helper.h"
#endif
Expand All @@ -50,13 +50,15 @@ static void AllReduce(phi::DenseTensor &tensor, // NOLINT

if (map->has(ring_id)) {
paddle::distributed::ProcessGroup *pg = map->get(ring_id);
auto pg_nccl = static_cast<distributed::ProcessGroupNCCL *>(pg);

std::vector<phi::DenseTensor> in_tensor;
std::vector<phi::DenseTensor> out_tensor;
in_tensor.push_back(tensor);
out_tensor.push_back(tensor);
paddle::distributed::AllreduceOptions opts;
opts.reduce_op = distributed::ReduceOp::SUM;
auto task = pg->AllReduce(in_tensor, out_tensor, opts);
auto task = pg_nccl->AllReduce(in_tensor, out_tensor, opts, true, true);
task->Wait();
} else {
auto dtype = platform::ToNCCLDataType(
Expand Down
6 changes: 4 additions & 2 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 @@ -23,7 +23,7 @@ limitations under the License. */
#include "paddle/phi/kernels/funcs/elementwise_functor.h"

#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
#include "paddle/fluid/distributed/collective/ProcessGroup.h"
#include "paddle/fluid/distributed/collective/ProcessGroupNCCL.h"
#include "paddle/fluid/platform/collective_helper.h"
#include "paddle/fluid/platform/device/gpu/nccl_helper.h"
#endif
Expand All @@ -43,13 +43,15 @@ static void AllReduce(phi::DenseTensor& tensor, // NOLINT

if (map->has(ring_id)) {
paddle::distributed::ProcessGroup* pg = map->get(ring_id);
auto pg_nccl = static_cast<distributed::ProcessGroupNCCL*>(pg);

std::vector<phi::DenseTensor> in_tensor;
std::vector<phi::DenseTensor> out_tensor;
in_tensor.push_back(tensor);
out_tensor.push_back(tensor);
paddle::distributed::AllreduceOptions opts;
opts.reduce_op = distributed::ReduceOp::SUM;
auto task = pg->AllReduce(in_tensor, out_tensor, opts);
auto task = pg_nccl->AllReduce(in_tensor, out_tensor, opts, true, true);
task->Wait();
} else {
auto dtype = platform::ToNCCLDataType(
Expand Down