Fix for fused Allgatherv

Signed-off-by: Yana Shchyokotova <yana.shchyokotova@intel.com>
horovod · May 17, 2019 · d49a3de · d49a3de
1 parent 0921a49
commit d49a3de
Show file tree

Hide file tree

Showing 2 changed files with 5 additions and 2 deletions.
diff --git a/horovod/common/ops/mlsl_operations.cc b/horovod/common/ops/mlsl_operations.cc
@@ -159,7 +159,7 @@ Status MLSLAllgather::Execute(std::vector<TensorTableEntry>& entries, const Resp
   }
 
   global_state_->timeline.ActivityStartAll(entries, MLSL_ALLGATHER);
-  auto mlsl_req = mlsl_context_->dist->AllGatherv((void*)sendbuf,
+  auto mlsl_req = mlsl_context_->dist->AllGatherv(sendbuf != nullptr ? (void*)sendbuf : buffer_data,
                                                   total_num_elements * element_size,
                                                   buffer_data, rcounts,
                                                   MLSL::DT_BYTE, MLSL::GT_DATA);

diff --git a/test/test_timeline.py b/test/test_timeline.py
@@ -44,7 +44,10 @@ def test_timeline(self):
                 hvd.init()
 
                 # Perform a simple allreduce operation
-                hvd.allreduce(torch.tensor([1, 2, 3], dtype=torch.float32), name='test_allreduce')
+                if 'MLSL_ROOT' in os.environ:
+                    hvd.allreduce(torch.tensor([1, 2, 3], dtype=torch.float32), name='test_allreduce')
+                else:
+                    hvd.allreduce(torch.tensor([1, 2, 3]), name='test_allreduce')
 
                 # Wait for it to register in the timeline.
                 time.sleep(0.1)