Skip to content

[Distributed] fix recreate nccl comm bug #73625

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

Merged
merged 1 commit into from
Jun 26, 2025
Merged

[Distributed] fix recreate nccl comm bug #73625

Show file tree
Hide file tree
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
2 changes: 1 addition & 1 deletion paddle/fluid/distributed/collective/process_group_nccl.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -988,8 +988,8 @@ void ProcessGroupNCCL::Restart() {
phi::distributed::P2POption p2p_opts = place_to_p2p_opts_.at(place_key);
phi::distributed::CommContextManager::RecreateNCCLComm(
store_, store_key, rank_, std::to_string(create_count_), &p2p_opts);
create_count_++;
}
create_count_++;
}

void ProcessGroupNCCL::SyncCalcStream(const Place& place,
Expand Down
3 changes: 2 additions & 1 deletion paddle/fluid/distributed/collective/process_group_nccl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
#pragma once

#include <chrono>
#include <map>
#include <memory>
#include <string>
#include <unordered_map>
Expand Down Expand Up @@ -276,7 +277,7 @@ class ProcessGroupNCCL final : public ProcessGroupWithStream {

uint64_t comm_seq_{0};
std::unordered_map<std::string, uint64_t> p2p_comm_seq_;
std::unordered_map<std::string, std::string> place_to_group_key_;
std::map<std::string, std::string> place_to_group_key_;

// TODO(sunyilun): attrs below will be removed later
std::mutex mutex_;
Expand Down
5 changes: 3 additions & 2 deletions paddle/phi/core/distributed/comm_context_manager.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -130,7 +130,7 @@ void CommContextManager::CreateNCCLCommContext(
void CommContextManager::RecreateNCCLComm(const std::shared_ptr<Store>& store,
const std::string& unique_comm_key,
int rank,
const std::string& hash_key,
const std::string& recreate_key,
const P2POption* p2p_opt) {
auto& comm_context_manager = CommContextManager::GetInstance();

Expand All @@ -139,7 +139,8 @@ void CommContextManager::RecreateNCCLComm(const std::shared_ptr<Store>& store,
PADDLE_ENFORCE_GPU_SUCCESS(phi::dynload::ncclGetUniqueId(&nccl_id));
}

std::string unique_key = "NCCLCommContext/" + unique_comm_key + hash_key;
std::string unique_key =
"NCCLCommContext/" + unique_comm_key + "/" + recreate_key;
if (rank == 0 || (p2p_opt && p2p_opt->is_p2p_op && p2p_opt->p2p_rank == 0)) {
std::vector<uint8_t> nccl_id_wrapper(
reinterpret_cast<uint8_t*>(&nccl_id),
Expand Down
Loading