diff --git a/projects/rccl/src/collectives/device/msccl_kernel_impl.h b/projects/rccl/src/collectives/device/msccl_kernel_impl.h index 3cbad93175..5e9b992e0e 100644 --- a/projects/rccl/src/collectives/device/msccl_kernel_impl.h +++ b/projects/rccl/src/collectives/device/msccl_kernel_impl.h @@ -372,7 +372,7 @@ __device__ __forceinline__ void mscclRunInterpreter( return; } if (t->hasDependence && tid == nthreads-1) - __atomic_store_n(&mscclFlags[bid].flag, (uint64_t) COMPUTE_FLAG(workIndex, iter, step), (t->type == MSCCL_REDUCE || t->type == MSCCL_RECV) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED); + __atomic_store_n(&mscclFlags[bid].flag, (uint64_t) COMPUTE_FLAG(workIndex, iter, step), ((t->type == MSCCL_REDUCE || t->type == MSCCL_RECV) && (t->dstBuffer != MSCCL_SCRATCH_BUFFER)) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED); step++; } }