Skip to content

Add device guard around MPI operations #22446

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

Closed
pietern wants to merge 1 commit into from
Closed

Add device guard around MPI operations #22446

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 11 additions & 0 deletions torch/lib/c10d/ProcessGroupMPI.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,8 @@

#include <map>

#include <c10/core/DeviceGuard.h>

#if defined(OPEN_MPI) && OPEN_MPI
#include <mpi-ext.h> // Needed for CUDA-aware check
#endif
Expand Down Expand Up @@ -316,6 +318,7 @@ std::shared_ptr<ProcessGroup::Work> ProcessGroupMPI::broadcast(
std::function<void(std::unique_ptr<WorkEntry>&)> runFunc =
[opts, this](std::unique_ptr<WorkEntry>& entry) {
auto data = (entry->src)[0];
c10::DeviceGuard guard(data.device());
std::unique_lock<std::mutex> globalLock(pgGlobalMutex_);
MPI_CHECK(MPI_Bcast(
data.data_ptr(),
Expand All @@ -337,6 +340,7 @@ std::shared_ptr<ProcessGroup::Work> ProcessGroupMPI::allreduce(
std::function<void(std::unique_ptr<WorkEntry>&)> runFunc =
[opts, this](std::unique_ptr<WorkEntry>& entry) {
auto data = (entry->src)[0];
c10::DeviceGuard guard(data.device());
std::unique_lock<std::mutex> globalLock(pgGlobalMutex_);
MPI_CHECK(MPI_Allreduce(
MPI_IN_PLACE,
Expand All @@ -363,6 +367,7 @@ std::shared_ptr<ProcessGroup::Work> ProcessGroupMPI::reduce(
void* sendbuf = (rank_ == opts.rootRank) ? MPI_IN_PLACE : dataPtr;
void* recvbuf = (rank_ == opts.rootRank) ? dataPtr : nullptr;

c10::DeviceGuard guard(data.device());
std::unique_lock<std::mutex> globalLock(pgGlobalMutex_);
MPI_CHECK(MPI_Reduce(
sendbuf,
Expand Down Expand Up @@ -402,6 +407,7 @@ std::shared_ptr<ProcessGroup::Work> ProcessGroupMPI::allgather(
std::vector<at::Tensor>& outputDataVec = entry->dst;
auto flatOutputTensor = newLikeFlat(outputDataVec);

c10::DeviceGuard guard(data.device());
std::unique_lock<std::mutex> globalLock(pgGlobalMutex_);
MPI_CHECK(MPI_Allgather(
data.data_ptr(),
Expand Down Expand Up @@ -456,6 +462,7 @@ std::shared_ptr<ProcessGroup::Work> ProcessGroupMPI::gather(
recvbuf = flatOutputTensor.data_ptr();
}

c10::DeviceGuard guard(data.device());
std::unique_lock<std::mutex> globalLock(pgGlobalMutex_);
MPI_CHECK(MPI_Gather(
data.data_ptr(),
Expand Down Expand Up @@ -529,6 +536,7 @@ std::shared_ptr<ProcessGroup::Work> ProcessGroupMPI::scatter(
}
}

c10::DeviceGuard guard(data.device());
std::unique_lock<std::mutex> globalLock(pgGlobalMutex_);
MPI_CHECK(MPI_Scatter(
sendbuf,
Expand Down Expand Up @@ -569,6 +577,7 @@ std::shared_ptr<ProcessGroup::Work> ProcessGroupMPI::send(
MPI_Request request = MPI_REQUEST_NULL;

{
c10::DeviceGuard guard(tensor.device());
std::unique_lock<std::mutex> globalLock(pgGlobalMutex_);
MPI_CHECK(MPI_Isend(
tensor.data_ptr(),
Expand All @@ -593,6 +602,7 @@ std::shared_ptr<ProcessGroup::Work> ProcessGroupMPI::recv(
MPI_Request request = MPI_REQUEST_NULL;

{
c10::DeviceGuard guard(tensor.device());
std::unique_lock<std::mutex> globalLock(pgGlobalMutex_);
MPI_CHECK(MPI_Irecv(
tensor.data_ptr(),
Expand All @@ -616,6 +626,7 @@ std::shared_ptr<ProcessGroup::Work> ProcessGroupMPI::recvAnysource(
MPI_Request request = MPI_REQUEST_NULL;

{
c10::DeviceGuard guard(tensor.device());
std::unique_lock<std::mutex> globalLock(pgGlobalMutex_);
MPI_CHECK(MPI_Irecv(
tensor.data_ptr(),
Expand Down