From 3e8b6bbcb25156650e0a1faa0d02119f9006f4d8 Mon Sep 17 00:00:00 2001
From: Baibaifan <39549453+Baibaifan@users.noreply.github.com>
Date: Tue, 19 Apr 2022 19:21:12 +0800
Subject: [PATCH] fix_nccl_barrier (#41970)

---
 .../collective/ProcessGroupNCCL.cc            | 19 +++----------------
 1 file changed, 3 insertions(+), 16 deletions(-)

diff --git a/paddle/fluid/distributed/collective/ProcessGroupNCCL.cc b/paddle/fluid/distributed/collective/ProcessGroupNCCL.cc
index 12de7d116e..86cc5b5db7 100644
--- a/paddle/fluid/distributed/collective/ProcessGroupNCCL.cc
+++ b/paddle/fluid/distributed/collective/ProcessGroupNCCL.cc
@@ -353,21 +353,8 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupNCCL::Broadcast(
 
 std::shared_ptr<ProcessGroup::Task> ProcessGroupNCCL::Barrier(
     const BarrierOptions& opts) {
-  std::vector<phi::GPUPlace> places;
-
-  if (!opts.place_ids.empty()) {
-    for (auto place_id : opts.place_ids) {
-      places.emplace_back(place_id);
-    }
-  } else if (!used_place_ids_.empty()) {
-    for (auto place_id : used_place_ids_) {
-      places.emplace_back(place_id);
-    }
-  } else {
-    auto numGPUs = GetSize();
-    int place_id = static_cast<int>(rank_ % numGPUs);
-    places.emplace_back(place_id);
-  }
+  // Only support single card single process
+  std::vector<phi::GPUPlace> places = {place_};
 
   std::vector<phi::DenseTensor> barrierTensors;
   barrierTensors.reserve(places.size());
@@ -375,7 +362,7 @@ std::shared_ptr<ProcessGroup::Task> ProcessGroupNCCL::Barrier(
   platform::CUDADeviceGuard gpuGuard;
   for (auto& place : places) {
     gpuGuard.SetDeviceIndex(place.GetDeviceId());
-    auto dt = full({1}, 0, phi::DataType::FLOAT32, phi::GPUPlace());
+    auto dt = full({1}, 0, phi::DataType::FLOAT32, place);
     barrierTensors.push_back(
         *std::dynamic_pointer_cast<phi::DenseTensor>(dt.impl()));
   }
-- 
GitLab