Remove memory fence in NVLink barrier.

This commit is contained in:
Shangyan Zhou
2025-06-24 17:51:28 +08:00
parent a15faa9ff0
commit f1d7a7c89f
2 changed files with 1 additions and 1 deletions

View File

@@ -446,7 +446,6 @@ barrier_block(int** barrier_signal_ptrs, int rank) {
// Add self-ranks, sub other ranks
if (thread_id < kNumRanks) {
atomicAdd_system(barrier_signal_ptrs[rank] + thread_id, FINISHED_SUM_TAG);
memory_fence();
atomicSub_system(barrier_signal_ptrs[thread_id] + rank, FINISHED_SUM_TAG);
}
EP_DEVICE_ASSERT(kNumRanks <= blockDim.x);